You are on page 1of 213

ÉCOLE DE TECHNOLOGIE SUPÉRIEURE

UNIVERSITÉ DU QUÉBEC

THÈSE PAR ARTICLES PRÉSENTÉE À


L’ÉCOLE DE TECHNOLOGIE SUPÉRIEURE

COMME EXIGENCE PARTIELLE


À L’OBTENTION DU
DOCTORAT EN GÉNIE
Ph. D.

PAR
Behnam EMAMI-MEHRGANI

OPTIMISATION DES POLITIQUES DE SURVEILLANCE PAR L'INTÉGRATION DU


CADENASSAGE/DÉCADENASSAGE DANS LA GESTION DE LA CAPACITÉ DE
PRODUCTION

MONTRÉAL, LE 15 MAI 2012

©Tous droits réservés, Behnam Emami-Mehrgani, 2012


©Tous droits réservés
Cette licence signifie qu’il est interdit de reproduire, d’enregistrer ou de diffuser en tout ou en partie, le
présent document. Le lecteur qui désire imprimer ou conserver sur un autre media une partie importante de
ce document, doit obligatoirement en demander l’autorisation à l’auteur.
PRÉSENTATION DU JURY

CETTE THÈSE A ÉTÉ ÉVALUÉE

PAR UN JURY COMPOSÉ DE :

Mme. Sylvie Nadeau, directrice de thèse


Département de génie mécanique à l’École de technologie supérieure

M. Jean Pierre Kenné, codirecteur de thèse


Département de génie mécanique à l’École de technologie supérieure

M. Marc Paquet, président du jury


Département de génie de la production automatisée à l’École de technologie supérieure

M. Anh Dung Ngô, membre du jury


Département de génie mécanique à l’École de technologie supérieure

M. Georges Abdul-Nour, membre du jury, examinateur externe


Département de génie industriel à l’Univérsité du Québec à Trois-Rivières

ELLE A FAIT L’OBJET D’UNE SOUTENANCE DEVANT JURY ET PUBLIC

LE 4 MAI 2012

À L’ÉCOLE DE TECHNOLOGIE SUPÉRIEURE


REMERCIEMENTS

Le travail présenté dans cette thèse a été effectué au sein de l’Équipe de recherche en sécurité
du travail (ÉREST) du département de génie mécanique à l’École de technologie supérieure
(ÉTS).

Je tiens particulièrement à remercier les professeurs Sylvie Nadeau et Jean Pierre Kenné,
directeurs de ce projet, pour m’avoir encadré tout au long de ce travail. Je les remercie du
temps qu’ils ont pu me consacrer, ainsi que de la confiance et du soutien qu’ils ont su
m’accorder pour mener à bien cette thèse.

Mes remerciements vont au professeur Marc Paquet qui m’a fait l’honneur de présider le jury
et aux professeurs Anh Dung Ngô et Georges Abdul-Nour qui ont accepté de participer au
jury de la soutenance.

J’adresse mes remerciements au Fonds de recherche du Québec - Nature et technologies


(FQRNT), à l’ÉTS, à l’ÉREST, à l'Association québécoise pour l’hygiène, la santé et la
sécurité du travail (AQHSST) et à la congrégation de Notre-Dame pour le soutien financier
qu’ils m’ont accordé tout au long de ce projet.

J’adresse également mes remerciements à tous mes collègues et amis (es) du laboratoire de
l’ÉREST-ÉTS qui m’ont apporté leur support dans la bonne humeur.

Pour mon père Hamid-Reza Emami-Mehrgani et ma mère Kobra Yousefi-Panah qu’ils


trouvent dans ce travail l’expression de mon amour profond. Enfin, je tiens aussi à remercier
tous ceux qui de près ou de loin m’ont soutenu durant la réalisation de cette thèse.
OPTIMISATION DES POLITIQUES DE SURVEILLANCE PAR L'INTÉGRATION
DU CADENASSAGE/DÉCADENASSAGE DANS LA GESTION DE LA CAPACITÉ
DE PRODUCTION

Behnam EMAMI-MEHRGANI

RÉSUMÉ

Ce travail vise à développer une stratégie optimale de gestion de la production dans le but de
minimiser les coûts de production tout en tenant compte de l’aspect sécurité par le biais du
cadenassage/décadenassage (C/D) durant les étapes de maintenance corrective et préventive.
Dans la littérature, des études qui traitent de l’intégration du C/D et de la gestion de
production sont presque inexistantes. De plus, dans le contexte actuel de l’industrie,
beaucoup de dirigeants et travailleurs pensent, à tort, que la planification et la réalisation des
différentes procédures de C/D prend beaucoup de temps. Par conséquent, ce temps de
production inactif est perçu comme diminuant la performance de l’entreprise vis-à-vis la
cadence de fabrication planifiée. Pour pallier à ce problème le concept novateur de :"mean
time to lockout/tagout (MTTLT)" a été développé. Le concept novateur du MTTLT consiste
à prendre en considération le temps de C/D en fonction des niveaux d'inventaires pour un
système manufacturier afin de trouver une politique optimale. Ce concept permet d'une part
de diminuer les coûts totaux de production, comprenant les coûts d’inventaire, de pénurie et
de maintenance sur un horizon infini et d'autre part d'augmenter le niveau de sécurité des
travailleurs. Cette thèse est présentée en trois (3) phases :

En premier lieu, le projet vise à intégrer le concept novateur du MTTLT dans la gestion de la
capacité de production sous forme de deux travaux. Tout d'abord, le MTTLT a été assimilé
pour un système en redondance passive (deux machines, non-identiques produisant un seul
type de pièce). Une modélisation a été faite par la chaîne de Markov homogène, et une
résolution numérique à travers des équations différentielles d'Hamilton-Jacoby-Bellman
(HJB) a conduit à la solution du système étudié. Nos résultats ont été validés par le biais
d’une analyse de sensibilité.

En deuxième lieu, l'influence de contrôle du MTTLT a été présenté pour une ligne de
production constituée de trois machines (deux machines sous forme redondance passive et
une troisième machine en série avec les précédentes) produisant un type de pièce. Une
modélisation a été faite par la chaîne de Markov homogène. Les outils utilisés à l'étape
précédente ont servi à résoudre ce cas. Le modèle de simulation a été également utilisé afin
de déterminer les coûts afférents à chaque combinaison obtenue par le biais du plan
d’expérience. Partant de ces coûts, une analyse de régression a été faite pour trouver la
nouvelle commande optimale du problème d’optimisation considéré. Et pour illustrer l’utilité
de nos résultats, une analyse de sensibilité a été effectuée.
VIII

En troisième lieu, ce travail s’est concentré sur la modélisation du MTTLT et l'erreur


humaine pour un sytème manufacturier flexible (FMS). La modélisation de l'erreur humaine
a été effectuée par une chaîne de Markov non-homogène, pour un système en redondance
passive produisant une seule pièce. La solution a été exposée par le biais des équations d'HJB
et nos résultats ont été confirmés par une analyse de sensibilité.

Mots clés: Systèmes manufacturiers, Mean time to lockout/tagout (MTTLT), Gestion de la


capacité, Redondance passive, Erreur humaine, Maintenance corrective, Maintenance
préventive.
OPTIMIZATION OF MONITORING POLICY BY INTEGRATION OF
LOCKOUT/TAGOUT IN PRODUCTION CAPACITY CONTROL

Behnam EMAMI-MEHRGANI

ABSTRACT

This work aims to develop an optimal strategy for production control in order to minimize
production costs while taking into account occupational safety aspects through lockout /
tagout during corrective and preventive maintenance activities. In the scientific literature,
studies that address the integration of the lockout/tagout and production control are almost
nonexistent. In addition, currently in industry, many managers and workers mistakenly
believe that planning and carrying out the various lockout / tagout procedures takes a long
time. Therefore, this idle production time is perceived as reducing the company's
performance towards the planned production rate. To overcome this problem the innovative
concept of "mean time to logout / tagout (MTTLT)" has been developed. The innovative
concept of MTTLT is to be considered the lockout / tagout time according to the inventories
levels for a manufacturing system to find an optimal policy. This concept allows on the one
hand to reduce total production costs including inventory, shortage and maintenance costs
over an infinite horizon and on the other hand to increase the safety worker level. This thesis
is presented in three stages:

In the first stage, the project aims to integrate the innovative concept of MTTLT into
production capacity control. The MTTLT has been applied to a passive redundancy system
(two machines, non-identical producing one type of part). A model has been presented using
a homogeneous Markov chain, and the numerical solution of the Hamilton-Jacoby-Bellman
(HJB) differential equations has provided the solution for our manufacturing system. Our
results have been validated through a sensitivity analysis.

In the second stage, MTTLT control has been applied to a production line consisting of three
machines (two machines as passive redundancy and a third machine in series with the
previous ones) producing one type of part. A model has been presented using an
homogeneous Markov chain. The same tools used in the previous stage have also been used
to solve this case. In this case, the simulation model has also been used to determine the costs
associated with each combination obtained through the experimental design. Based on these
costs, a regression analysis has been conducted to find the new optimal control of the
optimization problem under consideration. To illustrate the usefulness of our results, a
sensitivity analysis has been performed.

In the third stage, this work focused on MTTLT and human error modeling for a FMS.
Human error has been presented using a non-homogeneous Markov chain for a
manufacturing system with passive redundancy producing one type of part. The solution has
X

been obtained through the HJB equations and results have been confirmed by a sensitivity
analysis.

Keywords: Manufacturing systems, Mean time to lockout / tagout (MTTLT), Capacity


control, Passive redundancy, Human error, Corrective maintenance, Preventive maintenance.
TABLE DES MATIÈRES

Page

INTRODUCTION ........................................................................................................................1

CHAPITRE 1 REVUE CRITIQUE DE LA LITTÉRATURE...................................................3


1.1 Introduction ......................................................................................................................3
1.2 Commande optimale et l’environnement stochastique ....................................................4
1.3 Système manufacturier et système manufacturier flexible (FMS) ...................................6
1.3.1 Classification des systèmes manufacturiers flexibles (FMS) ....................... 7
1.3.1.1 Système manufacturiers flexibles (FMS) et la circulation du
flux ............................................................................................. 7
1.3.1.2 Dynamique d’un système manufacturier flexible (FMS) ......... 8
1.4 Définition de la redondance .............................................................................................8
1.4.1 Redondance active ........................................................................................ 9
1.4.1.1 Redondance active totale .......................................................... 9
1.4.1.2 Redondance active partielle ...................................................... 9
1.4.2 Redondance passive ..................................................................................... 9
1.5 Définition de la chaîne de Markov ...................................................................................9
1.5.1 Chaîne de Markov homogène et FMS .......................................................... 9
1.5.2 Chaîne de Markov non-homogène et FMS ................................................ 11
1.6 Taux de demande variable..............................................................................................11
1.7 Système de production complexe ...................................................................................12
1.8 Machines en série et FMS ..............................................................................................12
1.9 Machines en parallèle et FMS ........................................................................................13
1.10 Différents types de maintnenace ..................................................................................13
1.10.1 Maintenance préventive ............................................................................. 14
1.10.2 Maintenance corrective et curative............................................................. 14
1.11 Maintenance et accident au travail ...............................................................................17
1.12 Politique de cadenassage/décadenassage .....................................................................18

CHAPITRE 2 PROBLÉMATIQUE ET MÉTHODOLOGIE DE RECHERCHE....................21


2.1 Introduction ....................................................................................................................21
2.2 Buts de la recherche .......................................................................................................22
2.3 Questions de recherche...................................................................................................23
2.4 Méthodologie proposée ..................................................................................................23
2.5 Contribution et structure de la these ...............................................................................25

CHAPITRE 3 ARTICLE 1: LOCKOUT/TAGOUT AND PRODUCTION RISKS IN THE


PRODUCTION CONTROL OF MANUFACTURIGN SYSTEMS WITH
PASSIVE REDUNDANCY ...............................................................................31
Résumé ......................................................................................................................................31
XII

Abstract ......................................................................................................................................32
3.1 Introduction ....................................................................................................................32
3.2 Assumptions and Notations ............................................................................................34
3.2.1 Assumptions ............................................................................................... 34
3.2.2 Notations .................................................................................................... 35
3.3 Problem statement ..........................................................................................................37
3.4 Optimal conditions and numerical approach ..................................................................43
3. 5 Numerical example and sensitivity analysis ..................................................................46
3.6 Conclusion ......................................................................................................................53
Acknowledgments ......................................................................................................................54

APPENDIX 3.A .....................................................................................................................55

REFERENCES .....................................................................................................................59

CHAPITRE 4 ARTICLE 2: LOCKOUT/TAGOUT AND OPTIMAL PRODUCTION


CONTROL POLICIES IN FAILURE-PRONE NON-HOMOGENOUS
TRANSFER LINES WITH PASSIVE REDUNDANCY ..................................61
Résumé ......................................................................................................................................61
Abstract ......................................................................................................................................62
4.1 Introduction ....................................................................................................................62
4.2 Assumptions and notations.............................................................................................66
4.2.1 Assumptions ............................................................................................... 66
4.2.2 Notations .................................................................................................... 66
4.2 Problem statement ..........................................................................................................68
4.3 Numerical example and sensitivity analysis ..................................................................73
4.4 Control approach, experimental design and response surface methodology .................79
4.5 Conclusion ......................................................................................................................88
Acknowledgments ......................................................................................................................88

APPENDIX 4.A OPTIMAL CONDITIONS AND NUMERICAL APPROACH ....................90

APPENDIX 4.B .....................................................................................................................95

REFERENCES .....................................................................................................................97

CHAPITRE 5 ARTICLE 3: OPTIMAL LOCKOUT/TAGOUT, PREVENTIVE


MAINTENANCE, HUMAN ERROR AND PRODUCTION POLICIES OF
MANUFACTURING SYSTEMS WITH PASSIVE REDUNDANCY ............99
Résumé ......................................................................................................................................99
Abstract ....................................................................................................................................100
5.1 Introduction ..................................................................................................................101
5.2 Assumptions and notations...........................................................................................104
5.2.1 Assumptions ............................................................................................. 104
5.2.2 Notations ................................................................................................... 105
5.3 Problem statement ........................................................................................................106
XIII

5.4 Numerical example and results analysis ......................................................................111


5.5 Conclusion ....................................................................................................................123
Acknowledgments ....................................................................................................................123

APPENDIX 5.A OPTIMAL CONDITIONS AND NUMERICAL APPROACH ..................124

REFERENCES ...................................................................................................................127

CONCLUSION GÉNÉRALE ..................................................................................................129

TRAVAUX FUTURS ..............................................................................................................131

RÉFÉRENCES .........................................................................................................................133

ANNEXE 1 INTEGRATING LOCKOUT/TAGOUT WITH OPERATIONAL RISKS: THE


PASSIVE REDUNDANCY CASE ..................................................................137

ANNEXE 2 LOCKOUT/TAGOUT AND OPERATIONAL RISKS IN THE PRODUCTION


CONTROL OF A TRANSFER LINE WITH PASSIVE REDUNDANCY ....151

ANNEXE 3 LOCKOUT/TAGOUT AND HUMAN ERROR IN PRODUCTION CONTROL


OF MANUFACTURING SYSTEMS WITH PASSIVE REDUNDANCY ....171

ANNEXE 4 CADENASSAGE ET BAISSE DES COÛTS DE PRODUCTION .................179

ANNEXE 5 TEMPS MOYEN DE CADENASSAGE/DÉCADENASSAGE : UN OUTIL


D’OPTIMISATION DES POLITIQUES DE SURVEILLANCE ET DE
MAINTENANCE .............................................................................................186

ANNEXE 6 UTILISATION DES DIFFÉRENTS MODÈLES À UN CAS RÉEL .............189


LISTE DES TABLEAUX

Page

Table 3.1 State transition between the different modes ........................................................38

Table 3.2 Description of the system modes ..........................................................................38

Table 3.3 Mean transition rates between the different modes. .............................................40

Table 3.4 Parameters for the manufacturing systems (M1P1 and M2P1). ...........................48

Table 3.5 The variations of production threshold and average cost for M1P1 and M2P1. ..51

Table 4.1 Parameters of the numerical example ...................................................................74

Table 4.2 Polynomial coefficients.........................................................................................83

Table 4.3 Variations of optimal design factors based on inventory and backlog costs ........86

Table 5.1 Parameters of the numerical example .................................................................115


LISTE DES FIGURES

Page

Figure 1.1 Structure globale des systèmes manufacturiers .....................................................6

Figure 1.2 Organisation des travaux de maintenance dans un FMS .....................................16

Figure 1.3 Inégration du C/D dans un système manufacturier ..............................................20

Figure 2.1 Récapitulatif de la méthodologie de la recherche proposée .................................25

Figure 3.1 Displays the resulting transition diagram. ...........................................................39

Figure 4.1 Transfer line with passive redundancy producing one part type .........................69

Figure 4.2 State transition diagram .......................................................................................71

Figure 4.3 Production rate of M s and M 2 at mode 1 .............................................................74

Figure 4.3 Production rate of M s and M 2 at mode 3 .............................................................74

Figure 4.5 Production rate of M1 and M 2 at mode 7 ............................................................75

Figure 4.6 Production rate of M s at mode 2 ..........................................................................76

Figure 4.7 Production rate of M1 at mode 4 .........................................................................76

Figure 4.8 Production rate of M 1 at mode 8 ..........................................................................77

Figure 4.9 Production rate of M 2 at mode 5 ........................................................................78

Figure 4.10 Diagram of control approach ................................................................................79

Figure 4.11 Diagram of the simulation model .........................................................................81

Figure 4.12 Contour plot of the response surface ....................................................................84

Figure 4.13 Contour plot of the response surface ....................................................................84

Figure 4.14 Contour plot of the response surface ....................................................................85

Figure 4.15 Production threshold /Inventory cost....................................................................87


XVIII

Figure 4.16 Production threshold/Backlog cost .......................................................................87

Figure 5.1 Passive redundancy system producing one part type ..........................................107

Figure 5.2 State transition diagram ......................................................................................109

Figure 5.3 Age-dependent failure rate of the main machine ................................................114

Figure 5.4 Main machine production rate at mode 2............................................................117

Figure 5.5 Main machine production rate at mode 3............................................................117

Figure 5.6 Standby machine production rate at mode 5 .......................................................118

Figure 5.7 Standby machine preventive maintenance rate without human error at
mode 2 .................................................................................................................119

Figure 5.8 Standby machine preventive maintenance rate with human error at
mode 3 .................................................................................................................120

Figure 5.9 Trend of threshold value of main machine according to machine age ...............120

Figure 5.10 Trend of preventive maintenance rate of standby machine according to


machine age.........................................................................................................121
LISTE DES ABRÉVIATIONS, SIGLES ET ACRONYMES

x (⋅) Niveau d'inventaire

c+ Coût de mise en inventaire par unité de pièce et par unité de temps

c− Coût de rupture d'inventaire par unité de pièce et par unité de temps

cα Coût d'intervention sur la machine au mode α

cr1 Coût de réparation pour la défaillance de type 1

cr 2 Coût de réparation pour la défaillance de type 2

ctagout Coût de cadenassage/décadenassage

g (⋅) Coût instantané

J (⋅) Coût total actualisé

v (⋅) Fonction valeur

ρ Taux d'actualisation

d Taux de demande

ui (⋅) Taux de production i (i= 1,2)

uimax Taux de production maximal (i= 1,2)

hx Pas de discrétisation suivant l'inventaire du produit

ha Pas de discrétisation suivant l'âge de la machine

ξ (⋅) Processus stochastique

qαβ Taux de transition du mode α au mode β

min
v31 Taux de réparation minimal avec cadenassage/décadenassage pour la
défaillance de type 1 (une machine produisant un type de pièce)
XX

max
v31 Taux de réparation maximal avec cadenassage/décadenassage pour la
défaillance de type 1 (une machine produisant un type de pièce)

v31 (⋅) Taux de réparation avec cadenassage/décadenassage pour la défaillance de


type 1(une machine produisant un type de pièce)
min
v21 Taux de réparation minimal avec cadenassage/décadenassage pour la
défaillance de type 2 (une machine produisant un type de pièce)
max
v21 Taux de réparation maximal avec cadenassage/décadenassage pour la
défaillance de type 2 (une machine produisant un type de pièce)

v21 (⋅) Taux de réparation avec cadenassage/décadenassage pour le défaillance de


type 2 (une machine produisant un type de pièce)

ω51min Taux de réparation minimal avec cadenassage/décadenassage pour la


défaillance de type 1 de la machine principale

ω51max Taux de réparation maximal avec cadenassage/décadenassage pour la


défaillance de type 1 de la machine principale

ω51 (⋅) Taux de réparation avec cadenassage/décadenassage pour la défaillance de


type 1 de la machine principale

ω41min Taux de réparation minimal avec cadenassage/décadenassage pour la


défaillance de type 2 de la machine principale

ω41max Taux de réparation maximal avec cadenassage/décadenassage pour la


défaillance de type 2 de la machine principale

ω41 (⋅) Taux de réparation avec cadenassage/décadenassage pour la défaillance de


type 1 de la machine principale

x1 (⋅) Niveau d'inventaire en-cours

x2 (⋅) Niveau d'inventaire de produits finis

c1+ Coût de mise en inventaire d'en-cours par unité de pièce d'en-cours et par
unité de temps
XXI

c2+ Coût de mise en inventaire de produits finis par unité de pièce finie et par
unité de temps

c2− Coût de rupture d'inventaire de produits finis par unité de pièce finie et par
unité de temps

cr1 Coût de réparation de la machine principale M 1

cr2 Coût de réparation de la machine principale M 2

crs Coût de réparation de la machine de secours M s

u r1 Taux de réparation avec le cadenassage/décadenassage pour la machine

principale M 1

u r2 Taux de réparation avec le cadenassage/décadenassage pour la machine

principale M 2

urs Taux de réparation avec le cadenassage/décadenassage pour la machine de

secours M s

ui (⋅) Taux de production i ( i =1, 2, s )

uimax (⋅) Taux de production maximal i ( i =1, 2, s )

1,2
q12 Taux de défaillance des machines M1 et M 2

q12s Taux de défaillance de la machine de secours M s

crs Coût de réparation de la machine de secours S

u1 Taux de production de la machine principale M

u1max Taux de production maximal de la machine principale M

us Taux de production de la machine de secours S


XXII

usmax Taux de production maximal de la machine de secours S

24
w42 Taux de maintenance préventive sans erreur humaine

24 max
w42 Taux de maintenance préventive maximal sans erreur humaine

34
w42 Taux de maintenance préventive avec erreur humaine

34max
w42 Taux de maintenance préventive maximal avec erreur humaine

q14 Taux de réparation de la machine de secours S

q23 Taux de maintenance préventive sans erreur humaine au taux de maintenance

préventive avec erreur humaine de la machine de secours S

q24 Taux de maintenance préventive sans erreur humaine de la machine secours S

q34 Taux de maintenance préventive sans erreur humaine de la machine secours S

q54 Taux de réparation de la machine principale M

NIOSH National Institute for Occupational Safety and Health

NOHSC National Occupational Health and Safety Commission

AFIM Association Française de Normalisation

IRSST Institut de Recherche Robert-Sauvé en Santé et Sécurité du Travail

CSST Commission de la Santé et de la Sécurité au Travail

RSST Règlement sur la santé et la sécurité du travail

C/D Cadenassage/décadenassage

CSA Canadian Standards Association

SST Santé et Sécurité du Travail

FMS Flexible Manufacturing System

HJB Hamilton-Jacobi-Bellman
XXIII

HPP Hedging Point Policy

MTTLT Meant time to lockout/tagout

MTBF Mean time between failures

MTTR Mean time to repair

WIP Work-in-process

DPEDD Dynamic Programming Equations in Directional Derivative

DD Directional Derivatives

AQHSST Association québécoise pour l’hygiène, la santé et la sécurité du travail

GFA Gesellschaft für Arbeitswissenschaft

ASME American Society of Mechanical Engineers

FRSQ Fonds de la Recherche en Santé du Québec

FQRSC Fonds Québécois de la Recherche sur la Société et la Culture

FQRNT Fonds de recherche du Québec - Nature et technologies

RRSSTQ Réseau de Recherche en Santé et en Sécurité du Travail du Québec

ÉREST Équipe de recherche en sécurité du travail


INTRODUCTION

Aujourd’hui, l’objectif de l’industriel est de produire le maximum de biens, notamment en


minimisant le coût total de fabrication et en respectant les différentes obligations légales
auxquelles il est soumis. Cet objectif ne sera atteint qu’avec une bonne gestion de la
production, qui prendra en considération toutes les phases de la fabrication, en commençant
par la phase de conception jusqu’à la livraison du produit final.

La gestion de la production est efficace et fiable, si nous arrivons à optimiser les diverses
ressources de notre système manufacturier. L’environnement des systèmes manufacturiers est
stochastique. Parmis ces aspects stochastiques, nous retenons les risques d’accident, les
pannes des équipements et la variation de la cadence de fabrication. Afin de maîtriser les
phénomènes aléatoires des pannes et des accidents, nous devons trouver une solution,
permettant de diminuer leur fréquence ainsi que leur gravité. La nature incertaine de
l’environnement manufacturier découle des machines sujettes à des pannes et des réparations
aléatoires.

Au cours des dernières années, plusieurs travaux de recherche ont porté sur les problèmes
d’optimisation de la production des systèmes manufacturiers, en tenant compte de leur
complexité, la concurrence et des enjeux de la mondialisation des marchés, tels que Gharbi et
Kenné (2003), Charlot et al. (2006). Malgré tous ces efforts, les fréquences et les gravités des
accidents lors des interventions de maintenance restent encore assez élevées au Québec ainsi
que dans d’autres pays à travers le monde. Cette situation confirme les résultats publiés par
d’autres études au Québec et dans d’autres pays (Pâques 1989, Underwood 1992, NIOSH
1994, Windau 1998, ARIA 2000, Grusenmeyer 2000, NOHSC 2000, Mutawe 2002, AFIM
2004).

Afin d’obtenir une estimation globale de l’impact de la maintenance sur les risques pour les
travailleurs au Québec, une étude a été effectuée par l’Institut de Recherche Robert Sauvé en
Santé et Sécurité du Travail (IRSST) sur les accidents du travail mortels au Québec entre
2

1999 et 2003, (R-578 IRSST, 2008). Cette étude a été effectuée à partir des données
disponibles à la Commission de la Santé et de la Sécurité du Travail (CSST) pour les années
concernées. Entre 1999 et 2003, selon les rapports annuels de la CSST, les accidents du
travail ont entraîné 1275 décès. Ces rapports précisent que 163 d’entre eux (soit 13%) ont été
produits pendant une activité de maintenance : améliorer la machine, diagnostiquer, essayer,
dépanner, localiser une panne, inspecter, modifier, reconstruire, réparer et surveiller le
fonctionnement. À la lumière de ces résultats, il est envisageable de penser que les activités
de maintenance représentent au Québec une proportion significative des accidents mortels.

Ces aléas, lorsqu’ils se produisent peuvent mettre en danger les différents objectifs d’un
système manufacturier. Pour remédier à ce problème, certaines entreprises ont mis en place
une méthode de prévention de ces risques nommée cadenassage/décadenassage (C/D).
Malheureusement, encore jusqu’à aujourd’hui, certaines entreprises négligent les procédures
de C/D lors des différentes tâches de maintenance, ce qui explique les statistiques encore
élevées dans ce secteur d’activité. Aujourd’hui, la méthode de C/D est devenue une
obligation légale au Québec selon l’article 185 du RSST.
CHAPITRE 1

REVUE CRITIQUE DE LA LITTÉRATURE

1.1 Introduction

Bien que la littérature soit riche en travaux sur les systèmes manufacturiers flexibles (FMS),
peu d’entre eux traitent l’intégration de C/D en gestion de la production dans un
environnement stochastique. Aujourd’hui, le but dans le monde de l’industrie est
l’optimisation des systèmes manufacturiers, en répondant aux exigences des clients. Cette
optimisation n’est réalisable qu’en diminuant le coût total de production, tout en augmentant
la sécurité des divers travailleurs dans un environnement incertain.

Autrement dit, l’optimisation d’un FMS porte sur trois éléments fondamentaux : des facteurs
financés (contient le coût de production ainsi que le coût de la maintenance), des facteurs
techniques (la disponibilité, la maintenabilité et la fiabilité des équipements), des facteurs de
santé et de sécurité des travailleurs.

Dans la littérature, nous constatons que certains auteurs accordent plus d’importance aux
aspects techniques et économiques, tandis que d’autres mettent l’accent sur la sécurité des
travailleurs et des équipements dans un système manufacturier bien déterminé. La première
catégorie d’auteurs se focalise sur les paramètres qui permettent de contrôler et de déterminer
le taux de la production sur un horizon infini. La deuxième catégorie d’études se concentre
sur la santé et la sécurité des travailleurs en vue d’une prévention par l’élimination ou le
contrôle des aléas susceptibles de produire des accidents ou des maladies professionnelles.
Pour cette raison, nous exposons une vision globale de l’optimisation d’un FMS, tout en
augmentant la sécurité des travailleurs. Nous incorporons la politique de la maintenance
préventive, corrective avec le C/D dans la gestion de la capacité de production, ce qui mène à
une recommandation de commande optimale dans un environnement stochastique.
4

1.2 Commande optimale et l’environnement stochastique

Diverses méthodes ont été utilisées pour la recherche de solutions aux problèmes de
commande optimale reliés à la gestion de la production pour un FMS. Dans ces différentes
méthodes, les chercheurs ont utilisé divers dispositifs comme l’algorithme du Kushner
(Kushner et Dupuis, 1992), l’intelligence artificielle en se basant sur les algorithmes
génétiques (Basnet et Mize, 1986), l’heuristique (Thesen, 1999) et la simulation (Kenné et
Gharbi, 2004). Les résultats élaborés jusqu’ici comportent certaines lacunes: dans ces
travaux, certains paramètres comme les pannes, les interventions de maintenance, les
accidents et la demande n'ont pas été pris en considérations de façon aléatoires. La majorité
de ces méthodes ont été appropriées pour des systèmes purement déterministes. Pour cette
raison la théorie de la commande optimale dans un environnement stochastique a été mise au
point. Plusieurs chercheurs ont contribué à résoudre ce problème. Rishel (1975) a développé
les conditions d’optimum (nécessaires et suffisantes) pour obtenir la solution optimale en
utilisant la programmation dynamique. Older et Suri (1980) ont modélisé la commande
stochastique pour la planification de la production d’un FMS sujet aux pannes aléatoires
suivant un processus markovien homogène. Ils ont obtenu une équation de programmation
dynamique de la politique de commande optimale, mais sans obtenir la solution à cause de la
complexité du problème. Kimemia et Gershwin (1983) ont modélisé le système stochastique
par des processus Markovien homogènes avec un taux de transition constant pour déterminer
la politique de production dont le taux de production permet de minimiser le coût d'inventaire
et de pénurie. Gershwin (1983) a utilisé l’approximation du seuil critique et de la fonction
coût pour l’ordonnancement d’un FMS. La difficulté réside en la résolution des équations
d'HJB pour trouver la valeur du seuil critique. Akella et Kumar (1986) ont montré que pour
un système Markovien homogène (taux de transitions constants), la politique permettant de
maintenir un inventaire de sécurité non négatif pendant les périodes d’excès de capacité pour
prévenir les futures insuffisances de capacité est une politique de commande optimale. Cette
politique est appelée le seuil critique ou Hedging Point Policy (HPP). Dans le cas d’un
système manufacturier complexe, la fonction qui réalise le coût optimal appelé fonction
valeur doit satisfaire un ensemble d’équations différentielles appelées équation d’HJB,
(Boukas et Kenné, 1997). En sachant qu’il n’existe pas de solution analytique connue pour
5

les équations d’HJB, nous les résolvons par une approche numérique basée sur l’approche
numérique de Kushner (Kushner et Dupuis, 1992). Kenné et Gharbi (1999), ont traité le
problème de contrôle de flux optimal avec une distribution des pannes de la machine en
fonction de l’âge.

En se basant sur la revue de la littérature, nous constatons qu’il existe plusieurs types de
méthodes pour l’optimisation de la commande dans un environnement stochastique telles les
méthodes : markoviennes, semi-markoviennes, non-markoviennes, les heuristiques, le
principe du maximum de Pontryagin et la programmation linéaire. La méthode markovienne
peut être regroupée en deux (2) volets comme suit :

- Taux de transition d’un état à un autre de la machine sont constants: Modélisation par la
chaîne de Markov homogène;
- Taux de transition d’un état de la machine à un autre ne sont pas constants, mais
contrôlables: Modélisation par la chaîne de Markov non-homogène.

Lorsque le système est modélisé par une équation différentielle son avenir est uniquement
déterminé par sa situation présente, d’où le nom de dynamique déterministe. Étant donné que
l'environnement du système manufacturier est aléatoire ou stochastique, nous avons besoin
d'hypothèses pour le traduire. Dans cet environnement, plusieurs évolutions sont possibles à
partir de la situation présente, chacune d’elles ayant une certaine probabilité de se réaliser,
d'où la nécessité d'une modélisation par la chaîne de Markov (homogène ou non-homogène).
De plus la chaîne de Markov peut effectivement bien refléter le comportement dynamique
d'un système manufacturier. Avant de définir les processus de chaîne de Markov homogènes
et non-homogènes, pour mieux concevoir la suite de ce manuscrit, nous présentons la
structure des systèmes manufacturiers dans la Figure 1.1. Nous commençons par la suite la
classification des FMS ainsi que les définitions des systèmes en modes redondances.
6

Figure 1.1 Structure globale des systèmes manufacturiers


*MP : Matière première
*PF : Produit fini

1.3 Système manufacturier et système manufacturier flexible (FMS)

Avant de définir le système manufacturier flexible (FMS), nous commençons par la


définition d’un système manufacturier. Un système se définit comme un ensemble
d’éléments ou d’entités qui interagissent entre eux selon un certain nombre de principes ou
de règles. Si ces entités sont utilisées dans le but de la fabrication d’un bien, nous
construisons alors un système manufacturier. Le système manufacturier est un ensemble de
machines ou d’équipements, d’éléments de transport, d’unités d'emmagasinage, de
personnes, d’ordinateurs ou tout autre élément mis ensemble pour la fabrication (Gershwin,
2002). Un système manufacturier évolue au cours du temps selon les différents facteurs de
fabrication tels que les pannes, le nombre d’interventions (maintenance corrective,
préventive), l’âge des équipements et les variations de la demande. Tous ces facteurs ont un
impact direct sur le rendement d’un système manufacturier. Par conséquent, la variation de
ces facteurs dans un système manufacturier exige une flexibilité d’où la nécessité d’un FMS.
7

1.3.1 Classification des systèmes manufacturiers flexibles (FMS)

Un système manufacturier flexible (FMS) peut-être classifié par rapport à deux (2) critères :

- la circulation du flux (continus, discrets, hybrides);


- la dynamique d’un FMS (stochastique, déterministe, hybride).

1.3.1.1 Système manufacturiers flexibles (FMS) et la circulation du flux

Il est possible de classifier les FMS en trois (3) catégories selon la circulation de leur flux (le
flux est un transfert ou un échange d’une entité à une autre. Cela peut être un échange de
données, d’informations, d’équipements, de matières premières, etc.) :

1.3.1.1.1 Système manufacturier flexible continu

Lorsque le flux des matières ne s’arrête pas entre les postes successifs dans un environnement
manufacturier flexible, nous pouvons affirmer que le système manufacturier flexible est
continu.

1.3.1.1.2 Système manufacturier flexible discret

Nous parlons de système discret lorsqu’il y a des points de rupture dans le flux sous forme
d'en-cours et d'inventaires intermédiaires (Bironneau, 2000). Le mode discret d’un système
peut également se justifier par une suite d’opérations indépendantes réalisées sur des moyens
indépendants, par exemple deux (2) machines distinctes (Gharbi et Kenné,2003).

1.3.1.1.3 Système manufacturier flexible hybride

Les systèmes discrets et continus pris individuellement n’étant pas toujours en mesure de
satisfaire les exigences d’une entreprise qui se veut compétitive sur le marché, on peut
procéder à leur intégration dans un même système, donnant ainsi naissance aux systèmes de
fabrication hybrides (Bhattacharya et Coleman ,1994). Les systèmes manufacturiers flexibles
8

hybrides en général transforment une matière brute continue à travers des postes qui
fonctionnent de manière proportionnellement discrète.

1.3.1.2 Dynamique d’un système manufacturier flexible (FMS)

Il est possible de classifier les FMS en trois (3) catégories selon leur dynamique :

1.3.1.2.1 Dynamique stochastique

Un système manufacturier a une dynamique stochastique si au moins une de ses sorties ou un


de ses paramètres est aléatoire (Sader et Sorensen, 2003).

1.3.1.2.2 Dynamique déterministe

Un système manufacturier a une dynamique déterministe, si ses paramètres peuvent avoir


une dynamique évoluant de façon déterministe. (par exemple, la demande, lorsqu’elle est à
taux constant et connue)

1.3.1.2.3 Dynamique hybride

Lorsqu’il n’y a pas d’interaction entre les paramètres d’un système manufacturier flexible,
donc le système a une dynamique hybride (Cassandras et al., 1999).

1.4 Définition de la redondance

Il est inévitable qu'avec le temps les choses se détériorent, et cela, même quand elles sont au
repos. C'est la raison pour laquelle les entreprises organisent périodiquement le contrôle de
leurs équipements. Par ailleurs, nous ne sommes jamais à l'abri d'une défaillance soudaine et
imprévue d’où la nécessité des éléments en redondance. D’une manière générale les systèmes
en redondance sont divisés en deux (2) types :

 redondance active;
 redondance passive.
9

1.4.1 Redondance active

La redondance active consiste à utiliser plus d’éléments qu'il n'en faut strictement pour
remplir une fonction. Nous parlons de redondance active, quand tous les éléments
fonctionnent en permanence. Nous distinguons la redondance active totale et partielle.

1.4.1.1 Redondance active totale

Le système ne devient défaillant qu'avec la défaillance du dernier élément survivant. Par


définition, il s'agit d'un système dans lequel les éléments sont associés en parallèle.

1.4.1.2 Redondance active partielle

Nous parlons de redondance active partielle quand un système comporte n éléments, dont m
(m < n) strictement nécessaires pour qu’il fonctionne. Le système peut donc accepter (n-m)
défaillances.

1.4.2 Redondance passive

La redondance est qualifiée de passive quand les éléments abondants ne sont mis en service
qu'au moment du besoin; cela signifie que parmi n éléments seuls m sont en service. Ceci
implique que certains éléments seront en réserve ou en stock.

1.5 Définition de la chaîne de Markov

La chaîne de Markov contribue à spécifier le processus de production dans un environnement


stochastique pour un FMS, tout en tenant compte que l’état du système varie au cours du
temps. La chaîne de Markov peut avoir deux allures, homogène et non homogène.

1.5.1 Chaîne de Markov homogène et FMS

Une chaîne de Markov est homogène quand les taux de transition d’un état à un autre de la
machine sont considérés constants. Cela suppose que les équipements peuvent tomber en
panne même durant l’arrêt de production. Cette présomption est réaliste dans le cas où la
10

machine subit des facteurs exogènes. Il existe d’autres hypothèses pour la chaîne de Markov
homogène tels que :

- le temps moyen de réparation d’une panne est fixe;


- la disponibilité de chaque équipement composant le système est connue à l’avance.

La description des pannes et des réparations des équipements d’un FMS par le processus de
Markov homogène, a montré l’importance des pannes aléatoires sur le problème
d’optimisation des FMS. Il est à noter que ce processus ne pourrait pas être utilisé à cause de
la difficulté de la résolution des équations de programmation dynamique appelées équation
d’HJB. Pour résoudre ce problème, les chercheurs (Kimemia et Gershwin, 1983) posent
l’hypothèse que l’ensemble de la capacité d’un FMS est un hypercube. En décomposant les
équations différentielles couplées du type d’HJB sous forme d’équations normales, il est
possible de résoudre séparément afin de trouver les coûts de chaque état de la machine et
chaque pièce fabriquée (méthode de la décomposition et approximation quadratique). Dans le
même ordre d’idée, les experts précisent que pour un système constitué d’une machine
traitant un produit, la politique optimale est caractérisée par le niveau optimal d’inventaire ou
seuil critique (Akella et Kumar, 1986). Les experts ont développé une formule analytique
permettant de calculer le seuil critique, en supposant un taux de demande constant et que la
machine peut se trouver dans deux (2) états (opérationnel, panne). La politique de production
détermine le taux de production en fonction de l’inventaire. Cette méthode mène aux trois (3)
recommandations suivantes:

 produire exactement au taux de la demande, si le niveau d'inventaire et l’inventaire


optimal sont exactement les mêmes;
 ne pas produire si le niveau d'inventaire excède l’inventaire optimal;
 produire au taux maximal si l'inventaire est inférieur au niveau optimal d’inventaire.

Dans le cadre de l’optimisation d’un FMS, l’intégration du contrôle de C/D dans la gestion
de la production permet simultanément une vérification des taux de transition des
maintenances préventives ainsi qu’une vérification des modes opérationnels ou des modes
11

pannes. Donc, l’état de l’équipement est considéré comme un processus de la chaîne de


Markov non-homogène.

1.5.2 Chaîne de Markov non-homogène et FMS

Dans la majorité de cas, le FMS est modélisé par un processus de la chaîne de Markov non-
homogène. La probabilité qu’un équipement tombe en panne augmente avec son âge ainsi
que son cycle de fonctionnement. Le taux de transition du mode opérationnel à la
maintenance préventive est une variable de contrôle. La fiabilité de la machine est liée à son
âge (Boukas et Haurie, 1990). Les travaux de Boukas et Haurie (1990) et de Kenné et Gharbi
(1999) prennent en compte une probabilité de pannes des équipements croissante en fonction
de l’âge. Ces auteurs procèdent à une restauration de l’âge des équipements à zéro après la
maintenance préventive. La politique à seuil critique est obtenue pour contrôler le taux de
production en sachant que le taux de maintenance préventive est contrôlé par un âge critique
(l’âge optimal) dépendant de l’âge de la machine. L’âge critique est le point de transition du
mode opérationnel au mode maintenance préventive. Le contrôle de la maintenance
préventive permet d’améliorer la performance d’un FMS, tout en diminuant les coûts de
pénurie, les surplus et les réparations en restituant l’âge de la machine à zéro après
l’intervention. Gershwin et al. (1996) ont montré l’impact de la maintenance préventive sur
la minimisation des coûts de la production et l’assurance d’une meilleure qualité des pièces
fabriquées. En se basant sur la division de l’âge de la machine en quatre (4) segments (trois
modes opérationnels et un mode panne) et une qualité des pièces fabriquées en mode
opérationnel (bon, moyen et mauvais), Kenné (2003) a obtenu une loi de commande sous
optimale et a montré que la politique de la commande est asymptotiquement optimale.

1.6 Taux de demande variable

Dans la majorité de cas pour un FMS, les experts utilisent un taux de demande constant où
quand elle est variable, elle suit une distribution de la loi exponentielle. En sachant que le
taux de demande peut être variable et suit une distribution de Poisson, Feng et Yan (1999)
ont montré l’optimisation du seuil critique. Dans le même ordre d’idée, pour un système à
12

plus de deux (2) états, Sethi et al. (1992) ont résolu le problème par la technique de solution
de viscosité, ces derniers ont remarqué que la fonction valeur représentant le coût est
convexe. De plus, la fonction valeur est une solution de viscosité des équations d’HJB. En
effet, dans certains cas il arrive qu’il ne soit pas possible de représenter les pannes et les
réparations des équipements par des processus markoviens. Dans ce cas de figure, les
solutions ont été étudiées à partir des solutions obtenues par les processus markoviens, à
travers une extension du concept de la politique à seuil critique aux processus non-
markoviens. Kenné et Gharbi (2000), Kenné et Gharbi (2003), avec un processus non-
markoviens, en considérant une distribution des pannes et des réparations non exponentielles
ainsi qu’un taux de demande variable déterminent la solution optimale. Dans ce modèle
l’approche utilisée est basée sur les plans d’expérience et les techniques de simulation.

1.7 Système de production complexe

Bien que la littérature soit riche en travaux sur les FMS, peu d’entre eux traitent d’un FMS
avec plusieurs machines et plusieurs produits. Dans la plupart des modèles, pour simplifier la
résolution, nous considérons des cas simples, une machine et un produit. Dans le cas réel, le
FMS est généralement constitué de plusieurs machines. Il est à noter que plus le nombre des
machines augmente, plus la résolution du système est complexe, (Kenné et Gharbi, 2000;
Kenné et Gharbi, 2003). Les machines peuvent être disposées en série, en parallèle, en
redondance (active ou passive), ou même mixte série/parallèle, suivant les besoins de la
demande. La disposition des machines dans un FMS a un impact direct sur le coût total de
production. Pour les raisons susmentionnées, les auteurs supposent une série d’hypothèses de
façon à simplifier la résolution du problème d’optimisation de la production pour un FMS
complexe.

1.8 Machines en série et FMS

Dans un FMS, si les machines sont disposées en série le niveau d’inventaire intermédiaire
peut être considéré à capacité finie ou infinie. Si le niveau d’inventaire intermédiaire du
système a une capacité finie, la panne de la machine en aval peut entraîner le blocage de la
machine en amont ainsi que la panne de celle qui est en amont, l’arrêt de celle située en aval.
13

Sethi et al. (1992) ont exposé une politique de commande hiérarchisée sous optimale pour
deux (2) machines en considérant une capacité finie pour les tampons internes. Presman et al.
(1997) ont développé une politique sous optimale pour deux machines en série avec
l’approche proposée par Zhang et al. (1995). Kenné et Boukas (1998) ont montré une
amélioration de la disponibilité du système en introduisant le contrôle de la maintenance
corrective, pour un système constitué de plusieurs machines en série avec une capacité des
inventaires intermédiaires finies.

1.9 Machines en parallèle et FMS

La complexité due au FMS dont les machines sont disposées en parallèle est reliée à
l’ordonnancement quand le système produit plusieurs types de pièces. Dans le cas des
systèmes manufacturiers flexibles complexes, la fonction qui réalise le coût optimal appelée
fonction valeur doit répondre à un ensemble d’équations différentielles appelées équations
d'HJB, (Boukas et Haurie, 1990; Boukas et Kenné, 1997). Dans certains FMS complexes
dont les conditions d’optimum sont décrites par les équations d’HJB, la résolution des
équations d’HJB est insurmontable analytiquement et parfois même numériquement. Nous
référons le lecteur au travail de Kenné et Gharbi (2001). Une autre méthode proposée pour la
politique de production optimale pour un système constitué de plusieurs machines en
parallèle produisant plusieurs types de pièces est la résolution combinée. La résolution
combinée est basée sur les plans d’expérience, les modèles de simulation et la méthodologie
de surface de réponse, (Kenné et Gharbi 2000; Kenné et Gharbi 2003).

1.10 Différents types de maintenance

Plusieurs définitions sont attribuées au terme maintenance, notamment Monchy (1991) qui
définit la maintenance comme étant un ensemble de moyens de prévention, de correction ou
de rénovation suivant l’usage du matériel et suivant sa criticité économique, afin d’optimiser
le coût global de possession. La maintenance se décrit comme un ensemble d’opérations
d'entretien préventif et curatif destinées à accroître la fiabilité ou de pallier aux défaillances.
Supportant principalement la production, la maintenance joue un rôle important sur la
qualité, la sécurité et l’accroissement de la productivité dans toute entreprise moderne et
14

compétitive. Les différentes pratiques de maintenance ont évolué au fil des années en
fonction des différents besoins des sociétés industrielles.

Dans les paragraphes qui vont suivre, nous allons aborder la maintenance préventive et
corrective.

1.10.1 Maintenance préventive

Durant la Deuxième Guerre Mondiale, les choses ont gravement changé avec l’augmentation
de la demande des produits. Ce changement est du à la pression de la guerre alors que la
main d’œuvre baissait considérablement; ceci a entrainé une forte évolution au niveau du
développement des mécanismes industriels avec l’avènement des actifs numériques pouvant
pallier aux tâches multiples. En vue de cette évolution industrielle, Ben-Daya (2000) définit
cette période comme étant le commencement de la forte dépendance industrielle des actifs.
Les arrêts des actifs devenant plus considérables, les industries s’engagèrent à accorder plus
d’importance à l’aspect préventif, afin d’augmenter la disponibilité et la durée de vie des
équipements; d’où la naissance de la deuxième génération de la maintenance appelée
maintenance préventive réalisée tout simplement de façon systématique. Ce type de
maintenance regroupe la maintenance de type bloc et de type âge. La maintenance préventive
type bloc consiste à remplacer des équipements neufs, faits à des intervalles de temps fixes.
Ces remplacements ne tiennent pas compte de l’état du système et celui-ci, en cas de panne,
est remplacé par un système neuf (Barlow et al.,1996). Par contre, si la politique de
remplacement consiste à remplacer le système en cas de panne ou après une période de temps
de marche sans panne, nous sommes en face de ce qu’on appelle la maintenance de type âge.
L’âge de la machine est remis à zéro après chaque panne (Barlow et al.,1996).

1.10.2 Maintenance corrective et curative

Il y a soixante ans, les industries n’étaient pas bien mécanisées, les équipements étaient
surdimensionnés et la main d’œuvre était abondante. Ceci facilitait les réparations et rendait
moins coûteux les arrêts de production et la prévention des arrêts des équipements n’était pas
une grande priorité. Les tâches de maintenance étaient restreintes aux tâches d’entretien.
15

La maintenance était donc basée sur la réparation après défaillance. Ceci a donné naissance à
la première génération de maintenance appelée maintenance corrective. Une autre définition
a été proposée par Stevenson et Benedetti (2006) dans leur livre, la "Gestion des Opérations".
Ces auteurs ont distingué entre la maintenance corrective et la maintenance curative. Selon
ces auteurs, la maintenance curative consiste à intervenir sur le système pour sa remise en
fonctionnement en cas de panne, la maintenance corrective est plutôt une amélioration de
l’équipement et des installations en vue de diminuer le temps d’arrêt et les coûts de
maintenance.

L’organisation des maintenances corrective et préventive dans un FMS est présentée à la


figure ci-dessous :
16

EXIGENCES DE LA MISSION

Temps cumul d’immobilisation

Analyse de la maintenance corrective Analyse de la maintenance préventive

Durée Travaux

- Détection - Nature
- Diagnostic - Périodicité
- Remise en état - Durée
- Contrôle - Contrôle

Conception Politique de maintenance

Diminuer des temps pour :

Moyens en Moyens en Procédures


matériels personnels
Détection et diagnostic

Réparation

Contrôles

Autres éléments

Figure 1.2 Organisation des travaux de maintenance dans un FMS


17

Nous citons ici quelques études faites par les différents groupes de chercheurs en santé et
sécurité du travail (SST) à travers le monde sur l’importance du contrôle et de la gestion des
événements indésirables lors de la maintenance.

1.11 Maintenance et accident au travail

Les études faites par l'Occupational Safety and Health Administrations (OSHA) montrent
qu’environ 3 millions de travailleurs de maintenance d’équipements sont exposés à des
risques graves d’accident. En cas d’accident, le temps moyen perdu par travailleur est
d’environ 24 jours. D’autres études menées au Québec par l'IRSST (2008) précisent que près
du quart des accidents se produisent au moment des interventions des travailleurs sur les
équipements lors d’entretiens et de réparations. Cette situation ajoutée aux douleurs et aux
souffrances, liées aux risques ergonomiques (Harichaux et Libert, 2003), représente une perte
non négligeable pour les entreprises. D'autres études menées par Chinniah et Champoux
(2008) de l'IRSST, en 2005, précisent que les machines dangereuses ont causé la mort
d’environ 20 travailleurs au Québec et quelques 13 000 accidents ont pu être reliés aux
machines occasionnant des coûts de 70 millions de dollars pour la CSST. De plus, d’après les
statistiques de la CSST, au Québec, il y a en moyenne par année 6300 accidents relatifs aux
machines dont 17 décès par an. En 2008 seulement, nous avons dénombré 6 pertes en vies
humaines, 5225 accidents au cours de travaux d’installation, d’entretien ou de réparation de
machines. Pour remédier à ces problèmes, la méthode privilégiée est l’utilisation du C/D ou
consignation/déconsignation ou lockout/tagout. Par contre, dans le contexte actuel de
l’industrie, beaucoup de dirigeants et travailleurs pensent, à tort, que la planification et la
réalisation des différentes procédures de C/D prend beaucoup de temps. Par conséquent, ce
temps de production inactif est perçu comme diminuant la performance de l’entreprise vis-à-
vis la cadence de fabrication planifiée. (Charlot et al. 2006).

Nota : Nous avons cité ci-dessus, ces différents termes pour le C/D car le terme de C/D est
connu par les expressions consignation/déconsignation, lockout/tagout en dehors du Québec.
18

1.12 Politique de cadenassage/décadenassage

Un accident de travail peut être observé comme une perturbation, souvent avec des effets
beaucoup plus graves que la panne elle-même. Ces aléas, quand ils se produisent, font
beaucoup de dégâts, ce qui empêche l’entreprise de remplir sa mission. L’augmentation des
précautions de sécurité fait diminuer la fréquence des accidents, par contre, elles font
augmenter le coût total de production.

Parmi les solutions mises en place, pour répondre à ces critères, nous retenons le C/D. Cette
solution consiste à verrouiller une machine ou un équipement à l’aide d’un cadenas, puis, de
décharger toutes les sources d’énergie résiduelle (hydraulique, électrique, pneumatique et
etc.) de façon à éviter la mise en marche prématurée des équipements ou des machines
pendant qu’un technicien fait son intervention sur les appareils en question. La méthode de
C/D souvent est accompagnée par des fiches de signalisation affichées sur les tableaux de
commande et les points de démarrage de manière assez visible pour avertir les employés de
l’indisponibilité des appareils sous intervention. Par mesure de sécurité, le cadenas doit avoir
une seule clé et celle-ci doit être gardée par l’intervenant lui-même, dans le cas contraire, par
le responsable de l’équipe d’intervention. La norme canadienne CSA Z460-05 (2005) définit
le C/D ou consignation/déconsignation (Europe) comme «l’installation d’un cadenas ou
d’une étiquette sur un dispositif d’isolement des sources d’énergie conformément à une
procédure établie, indiquant que le dispositif d’isolement des sources d’énergie ne doit pas
être actionné avant le retrait du cadenas ou de l’étiquette conformément à une procédure
établie».

L’utilisation actuelle du C/D présente certaines lacunes. Un niveau de risque élevé existe
pendant l’intervention des machines en pannes, (Chinniah et Champoux, 2008). Le risque
d’erreur humaine au cours des interventions présente une probabilité importante d’impact
sous forme d’incidents ou d'accidents du travail chez les maintenanciers et sur la disponibilité
du système manufacturier. De plus l’unique utilisation du C/D sous forme procédurale est
une solution imparfaite, en effet des problèmes majeurs persistent : non planification des
procédures du C/D dans la gestion de la production. Par ailleurs limiter le seul usage du C/D
19

dans un but de maintenance corrective empêche l’utilisation du C/D en maintenance


préventive. Le manque de validation de faisabilité technique est un autre problème du C/D.
Afin de remédier à ces lacunes Matsuoka et Muraki (2001) ont proposé une méthode de
repérage assisté par ordinateur pour une usine de grande taille afin d'isoler les équipements
avec un nombre minimum d'opérations. L'objectif de ce travail a été d'améliorer
considérablement le temps de C/D en éliminant la redondance. Saunders et al. (2001) ont
préconisé le placardage pour offrir une alternative au C/D au niveau des équipements de
distribution de puissance. Charlot et al. (2006) ont montré que l'intégration du contrôle du
C/D à la planification de la production facilite le recours à l’implantation de mesures pour
diminuer les risques d’accident. Badiane (2010) a considéré la méthode de C/D comme une
activité qui peut être planifiée à part entière dans la planification de la production et non
conjointement avec l'activité de maintenance. Entre autres, ce travail a démontré que cette
façon de faire nous rend moins vulnérables aux inefficacités dans la planification de la
maintenance.

Nous avons discuté dans les paragraphes précédents de l’optimisation de la gestion de


production et des stratégies de maintenance des systèmes manufacturiers. Dans les cas
mentionnés ci-dessus, les auteurs ont peu abordé l’aspect de la sécurité des travailleurs. Pour
remédier à cette lacune et afin de diminuer les pertes en vies humaines, les arrêts maladie et
les coûts engendrés, nous nous sommes intéressés particulièrement à l’intégration du C/D
dans la gestion de la capacité de production. Nous avons également vérifié l'influence de
l'erreur humaine sur le C/D ainsi que les activités de maintenance (Figure 1.3).
20

Intégration du
cadenassage/décadenassage

MTTLT (Mean Time to


Lockout/tagout)

Gestion de la capacité de Gestion de la fiabilité


production humaine

Figure 1.3 Intégration du C/D dans un système manufacturier


21

CHAPITRE 2

PROBLÉMATIQUE ET MÉTHODOLOGIE DE RECHERCHE

2.1 Introduction

Aujourd’hui, la concurrence dans le monde industriel mène les entreprises à viser le


maximum de profits en minimisant les coûts liés à la fabrication. Ce but ne peut pas être
atteint sans optimisation des diverses ressources de l’entreprise. La compétitivité des
entreprises du XXI siècle repose sur trois (3) aspects; la quantité et la qualité des biens ainsi
que le respect de la santé et de la sécurité au travail.

Comme nous avons susmentionné, plusieurs travaux de recherche ont porté sur l’optimisation
de la gestion de production dans un système manufacturier. De nombreux auteurs s’accordent
sur le fait que de tels problèmes sont réalistes lorsqu’ils intègrent conjointement les aspects
qualité, inspection, santé et sécurité du travail, dégradation d’équipements, etc., (Gwo-Liang
et al., 2009).

L’environnement des systèmes manufacturiers est stochastique. Il est généralement parsemé


d’événements dont certains sont prévisibles, d’autres sont seulement contrôlables, d’autres
enfin qui ne sont ni prévisibles ni contrôlables (Gershwin, 2002). Ces aléas affectent la
gestion de production ainsi que la maintenance des machines et se manifestent à travers les
dégradations des machines, des biens, ce qui touchera la marge opérationnelle du système
manufacturier.

L’augmentation des pannes et des réparations associées à ces aléas aboutissent à la


dégradation des machines, la diminution de la disponibilité ainsi que l’augmentation des
risques liés à la maintenance corrective et préventive. Si ces éléments perturbateurs ne sont
pas pris en considération, la gestion de production ne permettra pas d’atteindre les trois
facteurs concurrentiels (la quantité et la qualité des biens et le respect de la santé et la
sécurité des travailleurs) susmentionnés.
22

Pour faire face aux pannes et aux accidents, les stratégies d’entretien et de contrôle sont
divisées en deux (2) volets : celles qui réduisent leurs fréquences et celles qui diminuent leurs
gravités (Nollet et al., 1994). Pour cette raison, jusqu’à aujourd’hui, plusieurs études faites
par les différents laboratoires de recherche sur les stratégies d’entretien et de contrôle, visent
à trouver des méthodes efficaces et fiables, permettant de maîtriser ces éléments de risque et
par la suite les rationaliser en production. (Charlot et al., 2006).

2.2 Buts de la recherche

Une fois que nous avons pris connaissance de ce qui est le C/D, nous élaborerons sur les
différents problèmes liés à cette méthode de prévention ainsi que les solutions afin de
l’améliorer. L’efficacité et l’efficience d’un FMS et sa rentabilité est liée à sa disponibilité et
sa maintenabilité (machines et équipements) ainsi qu’à la sécurité de ses travailleurs.

Au cours des dernières années, plusieurs travaux de recherche ont été effectués sur le C/D,
mais sans prendre en considération les points suivants :

1- La majorité des études proposent des modélisations analytiques, sans valider leur
faisabilité technique réelle, sans évaluer leur rentabilité;

2- Elles proposent différentes méthodes de C/D sous forme de procédure pour l’intervention
sur les équipements, mais non comme un outil fiable pour optimiser la fabrication ou la
production;

3- Elles élaborent et se centrent sur le C/D lors de la maintenance corrective;

4- Remplacer le C/D par d’autres mesures de réduction des risques, entre autres des
protecteurs verrouillés ou des dispositifs de sécurité, (R-587 IRSST, 2008).

Cette étude vise à intégrer le C/D dans la gestion de la capacité de production et également à
vérifier l'influence de l'erreur humaine sur la sécurité des travailleurs et le coût de production
durant les activités de C/D et de maintenance. Cette intégration et vérification a pour objectif
23

d'augmenter la sécurité des travailleurs, la disponibilité et la maintenabilité des machines,


tout en optimisant le coût total de production.

2.3 Questions de recherche

1- Comment, peut-on mettre en évidence le C/D dans la gestion de la capacité de production


sans la pénaliser?

2- Est-il possible de déterminer un modèle théorique qui sera proche de la réalité afin
d’intégrer le C/D dans la gestion de la capacité de production?

3- Est-ce qu’il existe un modèle mathématique adéquat qui pourra mettre à l’épreuve
l’intégration de l'erreur humaine durant les activités de C/D et de maintenance dans un
FMS?

2.4 Méthodologie proposée

En premier lieu, le projet a intégré le concept novateur du MTTLT dans la gestion de la


capacité de production sous forme de deux travaux. Tout d'abord, nous avons assimilé le
MTTLT pour un système en redondance passive (deux machines, non-identiques produisant
un seul type de pièce). Les variables de décision ont été les taux de production, le taux de
maintenance corrective incluant le C/D. Une modélisation a été faite par la chaîne de Markov
homogène (Akella et Kumar, 1986) et une résolution numérique à travers des équations
différentielles d'HJB a conduit à la solution du système considéré. L’objectif de ce travail a
été d’optimiser le coût de production, d’inventaire, de pénurie sur un horizon infini, tout en
augmentant le niveau de sécurité des travailleurs.

En second lieu, nous avons vérifié l'influence de contrôle du MTTLT pour une ligne de
production constituée de trois machines (deux machines sous forme redondance passive et
une troisième machine en série avec les précédentes) produisant un type de pièce. Les outils
de modélisation utilisés à l'étape précédente ont également servi à résoudre ce cas. Les
variables de décision ont été les taux de production des machines. Ces variables ont une
influence sur le niveau des inventaires intermédiaires et ceux des produits finis. Nous avons
24

développé les conditions d’optimum en utilisant une approche combinée, se basant sur une
combinaison de formalisme analytique, la simulation, le plan d'expérience et la méthodologie
de surface de réponse. Une analyse de sensibilité a illustré l'utilité de l'approche proposée.
Cette contribution a atteint son objectif en deux (2) volets : 1) trouver les variables de
décision permettant de réduire les coûts totaux production, comprenant les coûts d’inventaire
et de pénurie sur un horizon infini de planification. 2) libérer un espace-temps essentiel pour
minimiser les possibilités de contournement des dispositifs de protection ou d’escamotage
des procédures de C/D en intégrant la troisième machine sous forme redondance passive.

Et finalement, notre travail s’est concentré sur la modélisation du MTTLT et l'erreur humaine
pour un FMS. La modélisation de l'erreur humaine a été effectuée par une chaîne de Markov
non-homogène (Boukas et Haurie, 1990), pour un système en redondance passive produisant
une seule pièce. Les variables de décision ont été les taux de production des machines et le
taux de maintenance préventive avec ou sans erreurs. Les variables de décision ont une
influence sur le niveau des inventaires ainsi que la capacité du système manufacturier. Le
taux de défaillance du système manufacturier dépend de sa durée de vie, ce qui signifie que
la politique de maintenance préventive dépend de l'âge des machines. En effet, ce travail a
prévu une modélisation analytique par chaîne de Markov non homogène, dont la résolution
est faite d’une part par la résolution d’équations différentielles d’HJB et d'autre part par
analyse numérique. L’objectif de ce travail de recherche est composé de deux (2) volets : 1)
trouver les variables de décision permettant de réduire les coûts totaux de production,
incluant les coûts d’inventaire et de pénurie sur un horizon de planification infini. 2) vérifier
l'influence de l'erreur humaine sur le C/D ainsi que l'activité de maintenance préventive.
Le schéma global de méthodologie est présenté à la figure ci-dessous :
25

FV
Système manufacturier Formulation du Équation d’HJB
problème

RN

Évaluer la
performance de la loi
Loi de commande Approche de Kushner
de commande par la
optimale
simulation

Figure 2.1 Récapitulatif de la méthodologie de la recherche proposée


*FV : Fonction valeur
*RN : Résolution numérique
*HJB : Hamilton-Jacoby-Bellman

2.5 Contribution et structure de la these

Les résultats de cette thèse, utilisés dans des cas réels, montreront qu’une planification
adéquate du C/D (tel que dans la partie théorique) permet de diminuer le coût de production,
en améliorant la sécurité des travailleurs de maintenance et la disponibilité des machines. Les
travaux précédents de Charlot et al. (2006) ont permis de construire un modèle analytique de
l'intégration du contrôle du C/D à la planification de la production et de l'entretien en utilisant
une théorie de contrôle. Un protocole de prises de données en laboratoire et un protocole de
prises de données en entreprise ont été conçus par Malek et al.(2009). Le travail de Badiane
(2010) a permis de considérer la méthode de C/D comme une activité qui peut être planifiée à
part entière dans la planification de la production et non conjointement avec l’activité de
maintenance. De plus, ce travail a démontré que cette façon de faire nous rend moins
vulnérables aux inefficacités dans la planification de la maintenance. Ce projet de recherche
est une suite logique des travaux précédents et a pour objectif de modéliser le MTTLT dans
26

la gestion de la capacité de production. Il vérifie également l'influence de l'erreur humaine


sur la sécurité des travailleurs et le coût de production durant les activités de C/D et de
maintenance pour un FMS. Les résultats de ce projet de recherche permettront l’avancement
des connaissances, mais surtout la contribution à la réduction des risques de graves accidents
dans les milieux de travail et à l’optimisation du coût de production.

Cette thèse est constituée de cinq (5) chapitres comprenant les présents chapitres de la revue
critique de la littérature, problématique et méthodologie de recherche. Les trois (3) chapitres
qui forment le cœur du travail représentent des articles publiés ou soumis à des revues
scientifiques avec comité de lecture. Cette thèse propose également une approche de
formulation et de résolution du problème de commande optimale stochastique des systèmes
dynamiques en contexte manufacturier. Cette approche permet de trouver la stratégie
optimale de gestion de la capacité de production avec les activités de C/D et d'intégrer
l'erreur humaine dans la dynamique des systèmes manufacturiers en tenant compte du C/D.

Les contributions de cette thèse sont obtenues à travers la rédaction de trois (3) articles de
revues avec comité de pairs, la participation à sept (7) conférences avec comité de pairs et à
deux (2) conférences sans comité de pairs.

Les articles de revues sont référencés par :

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011 "Lockout/tagout and


operational risks in the production control of manufacturing systems with passive
redundancy", International journal of Production Economics 132 (2), pp.165-173.

- Emami-Mehrgani, B., Kenné, J.P. and Nadeau, S., 2012 "Lockout/tagout and Optimal
Production Control Policies in Failure-prone Non-homogeneous Transfer Lines With
Passive Redundancy",Accepted (January), International Journal of Production Research.
(Acceptance confimation: TPRS-2012-IJPR-0110).
27

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2012 "Optimal lockout/tagout,


preventive maintenance, human error and production policies of manufacturing systems
with passive redundancy", Submitted (March), Reliability Engineering & System Safety.
(Submission Confirmation: RESS-D-12-00096)

Les articles de conférences par comité de paires sont référencés par :

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2012 " Réduction des coûts de
production des systèmes manufacturiers sujets aux activités de
cadenassage/décadenassage", 34e Congrès de l’Association québécoise pour l’hygiène,
la santé et la sécurité du travail (AQHSST), 16 au 18 mai, Gatineau, Québec.

- Emami-Mehrgani, B., Kenné, J.P. and Nadeau, S., 2012 "Lockout/tagout and human
error in production control of manufacturing systems with passive redundancy",
Gesellschaft für Arbeitswissenschaft (GFA), 22 au 24 février, Kassel, Allemgane.

- Emami-Mehrgani, B., Nadeau, S. and Kenné, J.P. 2011 "Lockout/tagout and


operational risks in the production control of production line with passive redundancy",
International Mechanical Engineering Congress, American Society of Mechanical
Engineers (ASME), 11 au 17 Novembre, Denver, USA.

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011 "Optimisation des politiques de
surveillance par le biais du cadenassage/décadenssage", 33e Congrès de l’Association
québécoise pour l’hygiène, la santé et la sécurité du travail (AQHSST), 11 au 13 mai,
Trois Rivières, Québec.

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P. 2011. "Cadenassage et basse des
coûts de production", Travail et Santé 27(2),pp.13-15.
28

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P. 2011. "Temps moyen de


cadenassage/décadenassage : un outil d’optimisation des politiques de surveillance et de
maintenance", Travail et Santé 27(4),pp.22.

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P. 2010 "Integrating lockout/tagout with
operational risks: the passive redundancy case", Gesellschaft für Arbeitswissenschaft
(GFA), 24 au 26 Mars, Darmstadt, Allemagne, pp.475-482.

Les articles de conférences sans comité de paires sont référencés par:

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011 "Temps moyen de


cadenassage/décadenassage: un outil d’optimisation des politiques de surveillance",
ÉTS, 14 avril.

- Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011 "Optimisation des politiques de
surveillance par le bais du cadenassage/décadenassage ", ÉTS, 9 décembre.

Les trois articles de revues sont présentés dans les trois (3) prochains chapitres.

L’article du troisième (3) chapitre présente un modèle de détermination conjointe des


problèmes d’optimisation de coût, de sécurité des travailleurs, de maintenance corrective
pour deux types de défaillances d’un FMS sujet aux pannes et réparations. Le système
manufacturier est constitué de deux machines en redondance passive. Les variables de
décision sont le taux de production, le taux de maintenance corrective incluant le C/D. Cet
article a été publié dans la revue International Journal of production Economics sous la
référence : Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011 "Lockout/tagout and
operational risks in the production control of manufacturing systems with passive
redundancy", International journal of Production Economics 132 (2), pp.165-173.
29

Dans l’article du chapitre quatre (4), nous avons étudié une ligne de production constituée de
trois machines produisant un type de pièce (deux sous forme redondance passive et une
troisième en série avec les précédentes). Les machines sont sujettes à des pannes et à des
réparations aléatoires. Le problème de contrôle considéré est soumis à des contraintes non-
négatives des inventaires en-cours (work-in-process). Les variables de décision sont les taux
de production des machines. Les variables de décision influencent le niveau des inventaires
intermédiaires et des inventaires de produits finis. Cet article a été accepté dans la revue
International Journal of Production Research sous la référence : Emami-Mehrgani, B.,
Kenné, J.P. and Nadeau, S., 2011 "Lockout/tagout and Optimal Production Control Policies
in Failure-prone Non-homogeneous Transfer Lines With Passive Redundancy", Accepted on
January 26, 2012, Acceptance Confirmation : TPRS-2012-IJPR-0110.

Dans l'article du chapitre cinq (5), nous avons introduit la notion d'erreur humaine pendant
des activités de C/D et de maintenance préventive. Le système manufacturier considéré est
constitué de deux machines non-identiques sous forme redondance passive. Les machines
sont sujettes à des pannes, à des réparations aléatoires et à des activités de maintenance
préventive avec ou sans erreur humaine. Les variables de décision sont les taux de production
des machines et le taux de maintenance préventive avec ou sans erreurs. Les variables de
décision influencent le niveau des inventaires et la capacité du système manufacturier. Cet
artice a été soumis à la revue Reliability Engineering & System Safety sous la référence :
Emami-Mehrgani, B., Nadeau, S. and Kenné, J.P., 2012 "Optimal lockout/tagout,
preventive maintenance, human error and production policies of manufacturing systems with
passive redundancy", Submitted on March 7, 2012 , Submission Confirmation : RESS-D-12-
00096.

Pour terminer, nous récapitulons le bilan de ce travail et nous présentons les travaux futurs.
CHAPITRE 3

ARTICLE 1: LOCKOUT/TAGOUT AND PRODUCTION RISKS IN THE


PRODUCTION CONTROL OF MANUFACTURIGN SYSTEMS WITH
PASSIVE REDUNDANCY

Behnam Emami-Mehrgani1, Sylvie Nadeau1, Jean-Pierre Kenné1


1
Départment of Mechanical Engineering, École de technologie supérieure, Montreal
(Quebec) H3C 1K3, Canada
This chapter is published as an article in International Journal of production Economics.
2011. Vol.132, no 2, pp. 165-173

Résumé

Cet article présente un modèle de détermination conjointe des problèmes d’optimisation de


coût, de sécurité des travailleurs, de maintenance corrective pour deux types de défaillances
d’un FMS sujet aux pannes et réparations. Des décisions concernant la production ainsi que
la maintenance corrective doivent être prises afin de minimiser des coûts de production. Les
variables de décision sont le taux de production, le taux de maintenance corrective incluant le
cadenassage/décadenassage (C/D). L’objectif est d’optimiser les coûts de production,
d’inventaire et de pénurie sur un horizon infini par une planification efficace et efficiente du
C/D lors de la gestion de la production.

Pour atteindre nos objectifs, nous suivrons les étapes suivantes:

- vérification de l’influence de contrôle du C/D ainsi que le temps de réparation pour une
machine produisant un type de pièce;
- vérification de l’influence de contrôle du C/D pour un système en redondance passive
qui consiste en deux (2) machines (non-identiques) produisant un type de pièce.
32

Un exemple numérique et des analyses de sensibilité sont présentés pour illustrer l’utilité de
l’approche proposée.

Abstract

This paper addresses problems associated with production control and occupational safety in
a manufacturing system prone to failure involving two machines working in passive
redundancy. Machines turning out one part experience two modes of failure and repair:
firstly, where failure occurs when a machine remains in fair condition; and, secondly, where
such failure results in outright breakdown. Accordingly, we examine both modes of failure
for their impact on a flexible manufacturing system (FMS) with respect to production control
in terms of costs associated with lockout/tagout procedures and corrective maintenance. This
study seeks to identify optimal costs related to backlogs, inventories and maintenance over an
infinite planning horizon, along with levels of occupational risk where production control
includes efficient planning of lockouts/tagouts. Our study offers numerical methods which
may be employed to achieve optimal conditions in setting control policies. A numerical
example and sensitivity analysis support this approach.

Keyword: Flexible manufacturing systems; passive redundancy; corrective maintenance;

lockout/tagout.

3.1 Introduction

Industry strives to lower production costs and optimize profit through operating methods
consistent with various legal requirements. This target can only be achieved through the
application of controls governing all stages of production. We need to harness the various
resources available throughout a manufacturing system to obtain the benefits that flow from
reliable and efficient controls over output. Gharbi and Kenné (2003), and Charlot et al.
(2006), and others have produced studies recently addressing optimal production control in
flexible manufacturing systems. Despite these efforts, the severity and frequency of accidents
remains unacceptably high during maintenance procedures. Results published by a number of
33

agencies confirm these findings, including the National Occupational Health and Safety
Commission (NOHSC, 2000), Mutawe (2002) and the Association Francaise des Ingenieures
de Maintenance (AFIM, 2004). Lockouts/tagouts during such maintenance offer a potential
solution to this problem. The machine is padlocked. All sources of residual power (potential,
hydraulic, electrical, etc.) are discharged to avoid premature start-up during maintenance.
Many managers wrongly assume it takes too much time to plan and carry out a
lockout/tagout, fearing such downtime will lower productivity or reduce performance.
Researchers cited earlier focused on optimizing production control in manufacturing systems.
Many authors concur on the need for approaches that integrate quality control, occupational
analysis, health and safety, as well as wear and tear on equipment (Gwo-Liang et al., 2009).
It is possible to predict and control certain events while others occur randomly and are
beyond control within manufacturing systems (Gershwin, 2002). Control over production,
maintenance and profit margins are all adversely affected by events that damage equipment
or the goods produced in a manufacturing system. Reliability analysis and maintenance
policy optimization for cold-standby system has been studied by many researchers. Zhang
and Wang (2006), Zhang and Wang (2007) and Zhang et al. (2006) derived the expected
long-run cost per unit time for a repairable system consisting of two identical machines and
one repairman when a geometric process depicting working time is assumed. A two
machines cold-standby system is composed of a primary machine and a backup machine,
where the backup machine is only called upon when the primary machine fails. Cold-standby
systems are one of the most important structures in reliability engineering and have been
widely used in real manufacturing systems. Manufacturing systems operate in a stochastic
environment because machines break down and are repaired randomly. A rising number of
breakdowns and repairs degrades structure, reduces availability, and increases occupational
hazards associated with corrective maintenance. Left unattended, these disruptive elements
erode competitiveness expressed in terms of quantity, quality and occupational health and
safety. Maintenance strategies have been developed to address either the frequency or the
severity of accidents and breakdowns (Nollet et al., 1994). Occupational health and safety
researchers worldwide have carried out studies confirming the importance of monitoring and
controlling undesirable incidents during maintenance procedures. The Occupational, Safety
34

and Health Administration (OSHA) conducted studies concluding that some 3 million
workers risk serious injury in work performed to maintain equipment. Workers lose on
average about 24 days a year when an accident occurs. Nearly a quarter of all occupational
accidents occurring in Quebec result from work to maintain or repair equipment, according to
studies carried out by Institut de Recherche Robert Sauvé en Santé et Sécurité du Travail
(IRSST, 2008). Pain and suffering along with ergonomic hazards only add to the
considerable loss experienced by workers as well as firms in such circumstances (Harichaux
and Libert, 2003). An important consideration arises: How to optimize production yet ensure
occupational safety. The theory put forward makes certain reasonable assumptions in
answering this question (rates of demand for various products are held constant; machines
are flexible). Hence, we base this theory on the extension of stochastic optimal control theory
as in Akella and Kumar (1986). The control policy structure results from a value function
that solves the related Hamilton-Jacoby-Bellman (HJB) equations. We chose a numerical
approach to determine an approximate value function, rather than a true value function, to
construct the control policy. Given appropriate conditions, the control policy constructed
becomes asymptomatically optimal when variance from the true value function approaches
zero (see Kenné et al. (2003). Finally, we present a numerical example and sensitivity
analysis demonstrating the validity of this approach. Assumptions and notations are defined
in the next section. Section 3.3 states the problem. Section 3.4 offers the Hamilton-Jacoby-
Bellman (HJP) equations. We show how a numerical approach yields an approximate value
function. Section 3.5 provides a numerical example, sensitivity analysis and discussion of
results. Section 3.6 concludes the paper.

3.2 Assumptions and Notations

This paper incorporates the following assumptions and notations:

3.2.1 Assumptions

1- Costs of corrective maintenance for repair from an outright breakdown exceed costs of
corrective maintenance for repair of a machine in fair condition.
35

2- Mean corrective maintenance time under fair condition is briefer than mean corrective
maintenance time for a breakdown.

3- The corrective maintenance is carried out with lockout/tagout for two modes of failure.

4- The main machine is more robust than the standby machine.

5- The main machine returns to production immediately after each repair (corrective
maintenance for two modes of failure) and the standby machine stands idle.

6- Both machines do not fail simultaneously (with a good maintenance plan).

3.2.2 Notations

x (⋅) : inventory level

c+ : holding cost per unit of item over per unit of time

c− : backlog cost per unit of item over per unit of time

cα : cost incurred for the operation on the machine at mode α

cr1 : corrective maintenance cost of fair condition

cr 2 : corrective maintenance cost of outright breakdown

ctagout : lockout/tagout cost

g (⋅) : instantaneous cost

J (⋅) : total cost

v (⋅) : value function

ρ: discount rate

d: demand rate

ui (⋅) : production rate of the machine i (i= 1,2)

uimax : maximal production rate (i= 1,2)


36

h: discretization stepsize

ξ (⋅) : stochastic process

qαβ : transition rate from mode α to β

min
v31 : minimal corrective maintenance rate with lockout/tagout of fair condition (single
machine producing one part type)
max
v31
: maximal corrective maintenance rate with lockout/tagout of fair condition (single
machine producing one part type)

v31 (⋅) : corrective maintenance rate with lockout/tagout of fair condition (single machine
producing one part type)
min
v21
: minimal corrective maintenance rate with lockout/tagout of outright breakdown
(single machine producing one part type)
max
v21
: maximal corrective maintenance rate with lockout/tagout of outright breakdown
(single machine producing one part type)

v21 (⋅) : corrective maintenance rate with lockout/tagout of outright breakdown (single
machine producing one part type)

ω51min : minimal corrective maintenance rate with lockout/tagout for fair condition of the
main machine

ω51max : maximal corrective maintenance rate with lockout/tagout for fair condition of the
main machine

ω51 (⋅) : corrective maintenance rate with lockout/tagout for fair condition of the main
machine

ω41min : minimal corrective maintenance rate with lockout/tagout for outright breakdown of
the main machine
37

ω41max : maximal corrective maintenance rate with lockout/tagout for outright breakdown of
the main machine

ω41 (⋅) : corrective maintenance rate with lockout/tagout for outright breakdown of the main
machine

3.3 Problem statement

Many fielded systems use cold-standby redundancies (Sainaki, 1994; Pandey et al.,1996 and
Kumar et al.,1996) to achieve high reliability levels, since cold-standby redundancies are
believed to be more reliable that an analogous system with active redundancies (Subramanian
and Anantharaman, 1995; Ebeling, 1997 and Coit, 2001). This section develops a
manufacturing system for two machines that are not identical but producing one type of part
in passive redundancy. In the system, the main machine operates any time. The other
machine, named as the cold-standby redundancy, is in standby position and put online
immediately when the main machine fails. Then, the maintenance service is triggered to
restore the failed machine. After the failed machine is restored, it will be installed into the
standby position. The dynamics of the system are not different from those for one machine
producing one type of part. The state of the machine may therefore be characterized in terms
of a finite state Markov chain and the homogeneous Markov processes such as in Older and
Suri (1980). We consider three modes or sets of states for one machine producing one type of
part as follows:

1 if Operational;

ζ (t ) = 2 if Repair with lockout/tagout (corrective maintenance of outright breakdown);
3 if Repair with lockout/tagout (corrective maintenance of fair condition).

We do not present the model for single machine producing one type of part in this paper. The
reader referred to Charlot et al. (2006) for details of such a system. Thereafter, however, we
compare results obtained for a system in passive redundancy with one machine producing
one type of part. Hence, the dynamics for two machines in passive redundancy are in a
38

hybrid state comprised of a discreet state ζ (t ) = (ζ 1 (t ), ζ 2 (t )) and a continuous state x(t ) .The

discreet state ζ (t ) = (ζ 1 (t ), ζ 2 (t )) represents the machine’s state and continuous state x(t )

represents the stock level. We have ζ i (t ) =1 if machine at time-off, ζ i (t ) =2 if machine

operational, ζ i (t ) =3 if machine in repair with lockout/tagout (corrective maintenance of

outright breakdown) and ζ i (t ) =4 if machine in repair with lockout/tagout (corrective

maintenance of fair condition) with i = 1, 2. A positive value of x(t ) is inventory. A negative


value is backlog. This paper considers two machines operating at different rates of
production and failure. Consistent with assumptions set out in section 2.1, the Tables 3.1 and
3.2 display diverse modes for our system, ζ (t ) = (ζ 1 (t ) , ζ 2 (t )) ∈ M = {1, 2,3, 4,5} :

Table 3.1 State transition between the different modes

ζ 1 (t ) 2 2 2 3 4

ζ 2 (t ) 1 3 4 2 2

ζ (t ) 1 2 3 4 5

Table 3.2 Description of the system modes


Mode Machines state Description
1 (2,1) Main machine operational and standby machine idle
2 (2,3) Main machine operational and standby machine in repair with
lockout/tagout (corrective maintenance of outright breakdown)
3 (2,4) Main machine operational and standby machine in repair with
lockout/tagout (corrective maintenance of faire condition)
4 (3,2) Main machine in repair with lockout/tagout (corrective
maintenance of outright breakdown) and standby machine
operational
5 (4,2) Main machine in repair with lockout/tagout (corrective
maintenance of faire condition) and standby machine operational
39

Figure 3.1 Displays the resulting transition diagram.

Mean transition rates can be expressed in Table 3.3, from Table 3.1 and Figure 3.1:
40

Table 3.3 Mean transition rates between the different modes.

Transition rate From state To state


q12 (2, 1) (2, 3)
q13 (2, 1) (2, 4)
q14 (2, 1) (3, 2)
q15 (2, 1) (4, 2)
q21 (2, 3) (2, 1)
q24 (2, 3) (3, 2)
q25 (2, 3) (4, 2)
q31 (2, 4) (2, 1)
q34 (2, 4) (3, 2)
q35 (2, 4) (4, 2)
q41 (3, 2) (2, 1)
q42 (3, 2) (2, 3)
q43 (3, 2) (2, 4)
q51 (4, 2) (2, 1)
q52 (4, 2) (2, 3)
q53 (4, 2) (2, 4)

The 5 X 5 transition matrix Q of our system follows:

 q11 q
12
q
13
q
14
q 
15
 
 q21 q
22
0 q
24
q 
25
 
 q31 0 q q
33 34
q ,
35 (3.1)
ω q q q 0 
 41 42 43 44
ω q q 0 q55 
 51 52 53 

Hence, the transition matrix Q depends on:

ω41 : Corrective maintenance rate with lockout/tagout of outright breakdown for the main
machine;
41

ω51 : Corrective maintenance rate with lockout/tagout of fair condition for the main
machine;

Where ω41 = q 41 and ω51 = q 51 .

Thus, the considered stochastic optimal control problem may be characterized by the control
dependent transition matrix Q (⋅) .

qαβ (ω41 , ω51 ) ≥ 0 (α ≠ β ) (3.2)

qαα (ω41 , ω51 ) = −  qαβ (ω41 , ω51 ) , (α , β ) ∈ M (3.3)


α ≠β

Transition probabilities are expressed as:

qαβ (⋅)δt+0(δt) if α≠β,


p ξ (t +δ t) = β ξ (t) = α  =  (3.4)
1+q (⋅)δt+0(δt) if α=β.
 αβ

The set of admissible decisions at mode α (t ) and control policies (control variables) at mode

α (t ) : u1 (⋅), u2 (⋅), ω41 (⋅) and ω51 (⋅) is defined by:

((u1 (⋅), u2 (⋅), ω41 (⋅), ω51 (⋅)) ∈ R4 , 


 
Γ(α ) = 0 ≤ u1 (⋅) ≤ u1 ,0 ≤ u2 (⋅) ≤ u2 ,
max max
, (3.5)
ω min ≤ ω (⋅) ≤ ω max , ω min ≤ ω (⋅) ≤ ω max 
 41 41 41 51 51 51 

max
In equation (3.5), u
1
is the maximal production rate of the main machine, u2max is the

min and ω max are the minimal and


maximal production rate of the standby machine, ω41 41

maximal corrective maintenance rates of outright breakdown for the main machine and

ω51min and ω51max are the minimal and maximal corrective maintenance rates of fair condition
for the main machine.

We turn now to a more complete, though intuitive, definition of the controlled jump process
42

affecting the operating state of our system.

Let us rewrite equation (3.4) for controlled transition rates (i.e., transitions from modes 4 to 1
and 5 to 1) as follows:

q41 = ω41 (⋅) = lim 1/ dt ( P (ξ (t + δ t ) = 4 ξ (t ) = 1 , (3.6)


δ t →0

q51 = ω51 (⋅) = lim 1/ dt ( P (ξ (t + δ t ) = 5 ξ (t ) = 1 , (3.7)


δ t →0

The continuous component consists of a continuous variable x (⋅) corresponding to the


inventory/backlog of products. This state variable is described in the following differential
equation:

x ( t ) = u (⋅) − d (3.8)

Where :

u(⋅) = u1 if ξ (t ) = 1, 2,3 and u(⋅) = u2 if ξ (t ) = 4,5 with x(0) = x

Where d and x denote respectively the constant demand rate and the initial level of stock.

Objective of the control variable is to minimize discounted cost as follows:

∞ − ρ t
(
J α , x, u , u , ω , ω
1 2 41 51 )
= E e
0
(
g α , x, u1, u2 , ω , ω
41 51 ) 
dt x ( 0) = x, ξ ( 0) = α  ,

(3.9)

∀u ( ⋅) and u ( ⋅) ∈Γ (α ) ,
1 2

Where g ( x, u1 , u2 , ω41 , ω51 ) = c + x + (t ) + c − x − (t ) + cα , ∀α ∈ M is the instantaneous cost, c + ,

c − and cα , being the cost per unit to produce parts for inventory, backlog as well as

intervention cost on the machine.


With:
x + = m a x {0, x }
x − = m a x {− x, 0}
43

The system is considered feasible if:

 π ui
max
≥d (3.10)

knowing probability limits can be ascertained from the following equation for a system
conforming to a Markov process:

π (⋅)Q(⋅) = 0, (3.11)

π = 1
with:

π (⋅) : Limiting probabilities

Q (⋅) : Transition matrix rates

Hence, we have π (⋅) = (π1 , π 2 , π 3 , π 4 , π 5 ) representing the vector of limiting probabilities


from modes 1 to 5.
The problem is finding the production rates of machines and corrective maintenance rates for
both types of failure to minimize the total cost given by equation (3.9). The next section
treats this as a dynamic optimization problem.

3.4 Optimal conditions and numerical approach

Letν ( x,α ) denote the value function or minimum discounted cost for equations (3.9) as

expressed in the following equation:

ν ( x,α) =
u(α)∈Γ(α)
(
inf J α, x,u , u ,ω ,ω
1 2 41 51
) ∀α ∈M, (3.12)

With :

u (α ) = u1 if α = 1, 2,3 and u(α ) = u2 if α = 4,5


44

This value functionν ( x,α ) satisfies the set of Hamilton-Jacobi-Bellman (HJB) equations in
Kenné et al. (2003) and Charlot et al. (2006). We adapt the HJB equation to the optimal
control problem considered as follows:

ρν ( x,α) = { ∂
}
min ( u(α) − d) ν ( x,α) + g ( x,u,α) +  qαβν ( x,β ) , ∀α,β ∈M
u(α)∈Γ(α) ∂x α≠β
(3.13)

With:

u (α ) = u1 if α = 1, 2,3 and u(α ) = u2 if α = 4,5 .

Optimal control policy (u1* , u2* , ω41


*
, ω51
*
) denotes the minimizer over Γ(α ) on the right side
of equations (3.9). This policy is consistent with the value function obtained in equation
(3.12). Optimal control policy therefore rests in solving equation (3.13). Analytical solution
to (3.13) proves almost impossible, however.

Let us now expand the numerical method to solve for optimal conditions identified in the
previous section. This method is based on the Kushner approach found in Kushner and
Dupuis (1992), Boukas and Haurie (1990) and Kenné et al. (2003).

Kushner’s approach lets us use an approximation scheme for the gradient of value function
ν ( x,α ) . Let h denote the length of the finite difference interval of the variable x. Hence,

using h, ν ( x,α ) is approximated by ν h ( x,α ) and ν x ( x,α ) as follows:

1 h 
 h (ν ( x + h, α ) −ν ( x,α ) ) × (u (α ) − d ) if ( u (α ) − d ) ≥ 0 
h

ν x ( x,α ) × (u (α ) − d ) =  , (3.14)
 1 (ν h ( x,α ) −ν h ( x-h, α ) ) × (u (α ) − d ) otherwise 
 h 
With:

u (α ) = u1 if α = 1, 2,3 and u(α ) = u2 if α = 4,5 .

We manipulated the approximation arrived at in equation (3.14) to rewrite the HJB equations

(3.13) as follows:
45

 n u(α ) − d
−1
  n u(α ) − d j

ν ( x,α ) = min
h
h 
u∈Γ (α ) 
 ρ + qαα +  hj
j
 ⋅ 
  j =1 hj
(ν h ( x + h,α ) K +
 j =1
 

}
+ν h ( x − h,α ) K − )+g(x,α , ⋅) + Q.ν h ( x,α ) , (3.15)

With:

u (α ) = u1 if α = 1, 2,3 and u (α ) = u2 if α = 4,5 .

Where Γ (α ) is the discreet feasible control space or the so-called control grid and the other

term used in equation (3.15) is defined as:


1 if (u (α ) − d ) ≥ 0,
K+ = 
0 otherwise,

1 if (u (α ) − d ) < 0,
K− = 
0 otherwise,
With:

u (α ) = u1 if α = 1, 2,3 and u(α ) = u2 if α = 4,5 .

The equation presented in (3.15) can be interpreted as the infinite horizon dynamic
programming equation for a discreet-time, discreet-state decision process that addresses
problems confronted in optimizing output and controlling maintenance, as in Kenné et al.
(2003) and Charlot et al. (2006). The next theorem demonstrates that value function
ν h
( x, α ) approximates ν ( x, α ) for a small step size h.

Theorem

Let ν h ( x, α ) denote a solution to HJB equation (3.15). Assume they are constants C g and

K g as follows:

k
0 ≤ v h ( x, α ) ≤ Cg (1 + x g
)

Then,
46

lim v h ( x, α ) = v( x, α )
h→0

Proof

The proof of this theorem is given in Appendix 3.A.

The subsequent section uses a numerical example to illustrate the structure of the control
policies.

3. 5 Numerical example and sensitivity analysis

We have chosen not to include in this paper any review of the widely known homogenous
Markov process α ∈ M = [1, 2,3] used to consider a manufacturing system involving one

machine in three states. Markov's homogenous process α ∈ M = [1, 2,3, 4,5] is followed

subsequently to examine a system expanded to two machines that are not identical and
operate in five states with passive redundancy. The discreet dynamic programming equation
for one machine producing one type of part (M1P1) yields equations (3.16)-(3.18):

−1  u − d 
 u1 − d   1 (ν h ( x+h,1) k + +ν h ( x-h,1) k − ) 
ν ( x,1) = min
h
ρ + + q12 + q13   h , (3.16)
u1∈Γh (1)
 h  + g ( x,1) + q ν h ( x,2 ) + q ν h ( x,3) 
 12 13 

−1
 d  d 
ν ( x,2 ) = min
h
 ρ + + v21   (ν h ( x -h,2 ) + g ( x,2 ) + v21ν h ( x,1)  , (3.17)
u1∈Γ ( 2 ) 
h
h  h 

−1
 d  d h 
ν h ( x ,3 ) = min  ρ + + v31  h
 (ν ( x -h,3 ) + g ( x ,3 ) + v31ν ( x ,1)  , (3.18)
u1∈Γ  3  
h   h  h 
 

Equations (3.19)-(3.23) define a system pairing two machines that are not identical operating

in passive redundancy to produce the one part type:


47

 u1 − d h 
 (ν ( x+h,1) k + +ν h ( x-h,1) k − )
h
  
−1
 u1 − d
ν ( x,1) = min
h
 ρ + + q12 + q13 + q14 + q15   + g ( ) 12 ( ) 13 ( )  , (3.19)
x,1 + q ν h
x,2 + q ν h
x ,3
u1∈Γh (1)
 h   
+q14ν ( x,4) + q15ν ( x,5)
h h

 
 u −d 
 1 (ν h ( x+h,2) k + +ν h ( x-h,2) k − )
−1  
 u −d   h 
h
ν ( x,2) = min  ρ + 1 h h
+ q21 + q24 + q25  + g ( x,2) + q21ν ( x,1) + q24ν ( x,4) , (3.20)
u1∈Γh( 2)  h   
  h
+q25ν ( x,5) 
 
 

 u −d 
 1 (ν h ( x+h,3) k + +ν h ( x-h,3) k − )
−1  h 
 u − d   

ν h ( x,3) = min  ρ + 1 + q31 + q34 + q35  + g ( x,3) + q31ν h ( x,1) + q34ν h ( x,4)  , (3.21)
u1∈Γh  3  h  
 h 
+q35ν ( x,5) 
 
 

 u2 − d h 
 (ν ( x+h,4 ) k + +ν h ( x-h,4 ) k − ) 
h
  
−1
 u2 − d
ν ( x,4 ) = min
h
 ρ + + ω41 + q42 + q43   + g ( x,4 ) + ω ν
41
h
( x,1) + q42ν h
( x,2 )  ,(3.22)
u 2 ∈Γh ( 4)
 h   
+q43ν ( x,3)
h

 

 u2 − d h 
 (ν ( x +h,5) k + +ν h ( x-h,5) k − ) 
h
  
−1
 u2 − d
ν ( x,5) = min
h
 ρ + + ω51 + q52 + q53   + g ( x ,5 ) + ω ν
51
h
( x ,1) + q52ν h
( x ,2 )  , (3.23)
u 2 ∈Γh ( 5)
 h   
+ q53ν ( x,3)
h

 
48

We use the following computational domain:

Gxh = { x : −10 ≤ x ≤ 30} .

Note that the computational domain as Gxh = { x : −10 ≤ x ≤ 30} for a rather rapid

manufacturing system as our system (U = 0.4, d = 0.2), it allows us to better see the gap
between different values as well as different curves. Other parameters for the two cases
studied appear in Table 3.4.

Table 3.4 Parameters for the manufacturing systems (M1P1 and M2P1).
Parameter c+ c− ctagout cr1 cr2 u1max u2max d
Value 1 10 4 30 40 0.4 0.27 0.2
Parameter min
v21 max
v21 min
v31 max
v31 q12 q13 ρ h
(M1P1)
Value 0.1 0.07 0.125 0.2 0.006 0.008 0.001 0.25
(M1P1)
Parameter ω41min ω 41max ω51min ω51max q14 q15 ρ h
(M2P1)
Value 0.1 0.07 0.125 0.2 0.006 0.008 0.001 0.25
(M2P1)

The policy iteration technique solves Eqs. (3.16)–(3.23) for optimal conditions. Recall that
the control policies are obtained from the numerical resolution of the optimality conditions
given by Eqs. (3.16)–(3.23). The structure of production policies (u1 and u2) states that: if
the stock level is lower than a threshold level, then produce at maximum rates; else if the
stock level is upper than a threshold level produce nothing; otherwise produce at the demand
rate. Such structure in the control literature is called hedging point policy (HPP). We refer
the reader to Kimemia and Gershwin (1983) for the structure of this control policy and to
Akella and Kumar (1986) for the calculation of the threshold level of this control policy in
the case of producing one part type. The structure of corrective maintenance including
lockout/tagout policie (ω41 and ω51 ) recommend that: if the stock level is lower than a
threshold level, the corrective maintenance including lockout/tagout set at maximum rates,
49

otherwise the corrective maintenance including lockout/tagout set at minimum rates. The
joint optimization of production and corrective maintenance policies gives another version of
hedging point policy (HPP), which is an extension of so-called hedging point policy (HPP).

This control policy is treated in the equations (3.24)-(3.28):

uimax if x(⋅) < Z * ,



ui ( x, a) = d if x(⋅) = Z * , (3.24)
0 otherwise,

Where: u max (α ) = u1max if α = 1, 2,3 and u max (α ) = u2max if α = 4,5.

Where Z * is the optimal threshold value of stock level for each state of the machine.

Resulting equations are:


For one machine producing one part:

v21
max
if x (⋅) < A* ,
v21 ( x, 2) =  min (3.25)
v21 otherwise.

Where A* represents the optimal stock level at which the corrective maintenance rate with
min max
lockout/tagout needs to switch from v21 to v21 for outright breakdown.

v31
max
if x (⋅) < B* ,
v31 ( x,3) =  min (3.26)
v31 otherwise.

Where B* represents the optimal stock level at which the corrective maintenance rate with
lockout/tagout needs to switch from v31min to v31
max
for fair condition.
For two machines that are not identical and operating in passive redundancy to produce one

part:

ω 41
max
if x (⋅) < C * ,
ω 41 ( x, 4) =  min (3.27)
ω 41 otherwise.
50

Where C * is the optimal stock level at which the corrective maintenance rate with
lockout/tagout for outright breakdown on the main machine needs to switch from ω41min to

ω41max .

ω51max if x (⋅) < D * ,


ω51 ( x, 5) =  min (3.28)
ω51 otherwise.

Where D* is the optimal stock level at which the corrective maintenance rate with
lockout/tagout for fair condition on the main machine needs to switch from ω51min to ω51max .

Now, we present the influence of the set of control variables on the production threshold and
average cost according to the different values of backlog cost, holding cost and corrective
maintenance cost for single machine and two machines of passive redundancy in Table 3.5.
51

Table 3.5 The variations of production threshold and average cost for M1P1 and M2P1.

Maximal time

Maximal time

Maximal time

Maximal time
Minimal time

Minimal time

Average Cost

Minimal time

Average Cost

Average Cost

Average Cost
Minimal time
Z* (M1P1)

Z* (M1P1)

Z* (M2P1)

Z* (M2P1)

(M1P1)

(M1P1)

(M2P1)

(M2P1)
c− c+ Cr
1
Cr
2

10 5 30 40 3.25 1.5 1.5 0.75 1063.6 503.6 422.5 348.9


20 5 30 40 5.25 2.5 2 1.5 1627.4 743.1 626.1 467.6
30 5 30 40 6.5 3.5 3 2 2043.3 912.6 750.9 564.2
40 5 30 40 7 3.75 3.25 2 2351.5 1050.3 816.7 600.2
50 5 30 40 8 4.25 3.75 2.25 2591 1140.4 981.8 645.9
60 5 30 40 8.5 4.75 3.75 2.25 2809.7 1232.3 1047.8 692.5
70 5 30 40 9 5.25 4 2.5 3014.2 1316.6 1089.7 704.1
80 5 30 40 9.25 5.5 4 2.75 3219.6 1403.3 1173 765.2
90 5 30 40 9.75 5.5 4.25 2.75 3406.6 1478.8 1253.1 763.6
100 5 30 40 10.25 6 4.5 2.75 3599.3 1551.4 1371.2 811.9
30 1 30 40 3.25 1.25 1 0.75 1063.6 503.4 535.5 348.4
30 2 30 40 1.5 0.25 0.75 0.5 1384.1 629.7 545.5 363.8
30 3 30 40 0.5 0.25 0.5 0.25 1553.5 684.1 555.5 371.3
30 4 30 40 0.25 0.25 0.5 0.25 1672.3 764.2 567.5 380.2
30 5 30 40 0 0 0.5 0.25 1827 924.1 682.5 437.1
30 6 30 40 0 0 0.5 0.25 2128 1400.3 712.8 489.6
30 7 30 40 0 0 0.5 0.25 2727.1 1849.1 778.2 541.8
30 8 30 40 0 0 0.5 0.25 3331.3 2214.1 803.5 591.8
30 9 30 40 0 0 0.5 0.25 3832.1 2521.9 884.7 686.8
30 5 30 40 9.25 5.5 4.75 3.5 1063.6 503.6 315.1 239.7
30 5 40 40 9.25 5.5 4.75 3.5 1063.6 513.2 315.1 255.8
30 5 50 40 9.25 5.5 4.75 3.5 1075.4 522.8 328.3 275.7
30 5 60 40 9.25 5.5 4.75 3.5 1087.1 532.4 340.2 285.2
30 5 70 40 9.25 5.5 4.75 3.5 1098.9 542.1 351.7 294.7
30 5 80 40 9.25 5.5 4.75 3.5 1110.6 532.4 363.1 304.2
30 5 90 40 9.25 5.5 4.75 3.5 1122.4 561.3 374.5 314.2
30 5 100 40 9.25 5.5 4.75 3.5 1134.2 570.9 396.4 325.5
30 5 110 40 9.25 5.5 4.75 3.5 1145.9 580.5 414.4 341.5
30 5 120 40 9.25 5.5 4.75 3.5 1157.7 590.1 432.4 361.4
30 5 30 40 9.5 5.75 5 3.75 1063.6 503.6 326.3 256.5
30 5 30 50 9.5 5.75 5 3.75 1063.6 516.5 326.3 274.6
30 5 30 60 9.5 5.75 5 3.75 1078.5 529.5 341.6 297.5
30 5 30 70 9.5 5.75 5 3.75 1093.5 542.4 356.5 309.1
30 5 30 80 9.5 5.75 5 3.75 1108.4 555.4 372.9 322.54
30 5 30 90 9.5 5.75 5 3.75 1123.4 568.4 387.4 334
30 5 30 100 9.5 5.75 5 3.75 1138.3 581.5 402.8 345
30 5 30 110 9.5 5.75 5 3.75 1153.2 594.5 426.7 359.3
30 5 30 120 9.5 5.75 5 3.75 1168.2 607.4 448.6 377.3
30 5 30 130 9.5 5.75 5 3.75 1183.1 620.2 469.6 400.3
52

The first section of Table 3.5 presents the variations in stock levels based on backlog costs
for one machine as well as two dissimilar machines operating in passive redundancy to make
one type of part. In this division, the lockout/tagout and corrective maintenance rates are set
at minimum and maximum values for fair condition and outright breakdown. Note that
varying backlog costs have little impact on stock levels if we increase lockout/ tagout and
corrective maintenance rates. Moreover, variations in stock levels drop rapidly when a
standby machine is added to our system. As this division shows, control over these variations
is enhanced if we maximize lockout/tagout and corrective maintenance rates for both modes
of failure. Average costs based on variations in backlog costs appear also in this division of
Table 3.5.

It shows that the variations in average cost are directly related to stock levels when
lockout/tagout and corrective maintenance rates are set at minimum and maximum values for
fair condition and outright breakdown. We can improve this situation by adding the standby
machine at maximal lockout/tagout and corrective maintenance rates for both modes of
failure.

The second section of Table 3.5 shows the variations in stock levels based on holding costs
for one machine as well as two dissimilar machines operating in passive redundancy to make
one type of part. This section discloses significant differences in the two systems under
study. Stocks drop precipitously as holding costs rise when a lone machine turns out one type
of part. We can mitigate this variation by maximizing lockout/tagout as well as corrective
maintenance rates for either mode of failure. A far more homogenous variation occurs with
passive redundancy, because it makes possible to respond to demand at all times.

The same section of Table 3.5 shows that how average costs and holding costs move in
tandem. Passive redundancy improves control over such costs; lockout/tagout and corrective
maintenance rates are maximized to prevent both modes of failure. In the third section of
Table 3.5, we illustrate the variations in stock levels and the corrective maintenance costs for
fair condition occurring in either of the two cases under study. Increased corrective
53

maintenance costs produce no material variation in stock levels as well as average costs for
fair condition. The variations in stock levels and average costs are far lower under passive
redundancy than other cases where lockout/tagout and corrective maintenance rates are set at
their highest levels to counter either mode of failure.

The last section of Table 3.5 confirm that when corrective maintenance costs rise in response
to outright breakdown, the variations in stock levels and average costs are far lower in the
passive redundancy than in other cases, where lockout/tagout and corrective maintenance
rates are maximized to combat failure modes 1 and 2. These results support the analysis of
the previous section.

Optimal costs of production and corrective maintenance including lockout/tagout procedures


for both modes of failure may be determined using the analytical model presented in this
paper. The numerical approach demonstrates conclusively that the resulting policy is optimal
and enhances machine availability. Control policy for our systems considers extension of the
hedging point structure. Without limiting in any way the generality of this proposal, this
model is based on certain assumptions relating to a pair of machines which are not identical
and which operate in passive redundancy. Given certain conditions, extended versions of this
model might be adopted across a number of industrial sectors.

3.6 Conclusion

Purpose of this paper is two-fold. We first sought to verify the influence of lockout/tagout
and corrective maintenance rates on two modes of failure in a manufacturing system
consisting of one machine producing one type of part. Secondly, we added a standby
machine which differed from the main machine to monitor the influence of passive
redundancy within our system. This work demonstrates clearly that passive redundancy
optimizes production and maintenance costs while enhancing occupational safety. Even
greater benefits accrue if effective lockout/tagout and maintenance planning occurs in
concert with production control. Subsequent research will seek to verify the influence of
lockout/tagout and preventive maintenance rates on a production line operating in a flowshop
54

(such as an assembly line) with internal buffers for two machines working along side a
standby machine.

Acknowledgments

We gratefully acknowledge for financial support of this work from the following agencies:
Réseau de Recherche en Santé et en Sécurité du Travail du Québec (RRSSTQ); Fonds de la
Recherche en Santé – Québec (FRSQ); Fonds Québécois de la Recherche sur la Société et la
Culture (FQRSC); Fonds Québécois de la Recherche sur la Nature et les Technologies
(FQRNT) and Institut de Recherche Robert-Sauvé en Santé et en Sécurité du Travail
(IRSST).
55

APPENDIX 3.A

Proof

Let us consider :

- z( x, α ) = x + b( x, α )Δt h ( x, α ) as the point which is reachable from x under control α in

direction b ( x, α ) in time Δt h ( x, α ) ;

- uαh ( x ) (uα = u1 if α = 1, 2, 3 and uα = u 2 if α = 4, 5) minimizing value of α in equation

(3.15);
- {ξ h
n , n < ∞} the approximating controlled of our system;

- {u h
n , n < ∞} the actual random variables which are the optimal control actions.

We define Δtnh = Δt h (ξ nh , unh ) and tnh =  i =0 Δtnh .


n −1

Then, we have:

E ξ nh+1 − ξ nh ξ nh = x, unh = α  = b( x, α )Δt h ( x, α ) = O ( h), (3.15.1)

cov ξ nh+1 − ξ nh ξ nh = x, unh = α  = O (h 2 ), (3.15.2)

Let Enh denote the expectation, given the state and control actions including time n.

We rewrite:
ξ h (t ) = x +  Δξnh and Δξ nh = ξ nh+1 − ξ nh in form Δξ nh = Enh Δξ nh + ( Δξ nh − Enh Δξ nh )
n: tnh+1 ≤t

Then, we have:
ξ h (t ) = x +  Enh Δξnh +  (Δξnh − Enh Δξnh ), (3.15.3)
n: tnh+1 ≤t n: tnh+1 ≤t

By knowing that the right hand sum in (3.15.3) is a continuous time interpolation of a
martingale and the variance of (3.15.2) is:
56

E  Δtnh O (h) = O (h)t


n: tnh+1 ≤t

Hence, we have:
2

E sup
t ≤T
 ( Δξ − E Δξ )
h
n
h
n
h
n = O ( h) for any t < ∞
n: tnh+1 ≤ t

Thus the effects of that right hand term in (3.15.3) disappear in the limit.
We write the right hand sum in (3.15.3) simply as O(h). We define the continuous parameter
interpolation uαh (⋅) by uα (t ) = unh (uα = u1 if α = 1, 2,3 and uα = u2 if α = 4,5) on the

interval tnh+1 , tnh ) .

Now, using (3.15.1), we have:


ξ h (t ) = x +  Enh Δξnh + O(h)
n: tnh+1 ≤t
t
= x +  b(ξ h (s), uαh ( s))ds + O(h), (3.15.4)
0

with:
uα = u1 if α = 1, 2,3 and uα = u2 if α = 4,5.
We now proceed to show that there is some admissible control such that the paths of ξ h (⋅)
are actually good approximations to a solution of (3.15). Because Δtnh = O (h) and Δtnh ≥ k1h,

the piecewise linear interpolations of the paths of process ξ h (⋅) are equicontinuous in

( w and h ) . Thus, for each fixed value of the probability space variable w , each subsequence

of {ξ h (⋅)} has a further subsequence which converges to some limit uniformly on each

bounded time interval. We now fix the sample space variable w and let hn (w) index a
convergent subsequence of the piecewise linear interpolations of

{ξ h
(⋅), ui h (⋅)} with limit denoted by x(⋅, w), uα (⋅, w) ( uα = u1 if α = 1, 2,3 and uα = u2 if α = 4,5).
By the convergence and the compactness of Uα we have uα (t , w) ∈Uα (uα = u1 if α = 1, 2,3
and uα = u2 if α = 4,5).
The uniform convergence implies that :
t
x(t , w) = x +  b( x( s, w), uα ( s, w))ds, (3.15.5)
0

with:
uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5.
57

Thus, the limit path satisfies (3.15) with an admissible control. Also, it is easy seen that
V hn (w) (x,α )converges to J (x, u1(w), u2 (w),α,ω 41 ,ω 51).

Due to the minimality of V ( x, α ) , we have:

J ( x, u1 (w), u2 (w), α , ω 41 , ω 51 ) ≥ V ( x, α ), (3.15.6)

Because this holds for each w ,


limV h ( x, α ) ≥ V ( x, α ), (3.15.7)
h→0

We now need to get the reverse inequality to (3.15.7). Namely, that there exists an optimal
admissible control for finite difference type approximations and that the control can be
arbitrarily well approximated by a control which is piecewise constant. That is, given any
ε > 0, there is an − optimal control uαε (⋅) (uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5) of the

following form: there is δ > 0 and a finite number of points


U αε inU α (uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5) such that uαε (⋅) is U αε -valued and is

constant on the intervals [ jδ , jδ + δ ) . we now apply this ε -optimal control to a passive

redundancy system and use the minimality of V h ( x, α ) for the controlled our system to get
the reverse inequality to (3.15.7).
We proceed as follows: we fix ε >0 and we define a sequence
uαh ,ε (⋅) (uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5) of controls for the passive redundancy

system by adapting the above ε -optimal control


uαε (⋅) (uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5) in the following natural way:

Let us consider h small enough such that δ > sup x ,α Δt h ( x, α ). We now define the sequences
n −1
unh ,ε , Δtnh = Δt h (ξ nh ,ε , unh ,ε ), and tnh =  Δtih , recursively by unh ,ε = uαε ( jδ ) for all n such that
0

tnh ∈ [ jδ , jδ + δ ) , for each j.

Let us ξ h ,ε (⋅) and uαh ,ε (uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5) denote the continuous

parameter interpolation (interpolation intervals Δtnh ). Note that uαh ,ε converges to


58

uαε (⋅) as h → 0, except possibly at the points jδ . Now, we fix w and choose a convergent

subsequence of ξ h,ε (⋅) (the sequence need not be the same for each w ). Let us xε (⋅, w) denote
the limit. Then, following the analysis which let to (3.15.5), we get that:
t
xε (t , w) = x +  b( xε ( s, w), uαε ( s, w))ds
0

with:
uα = u1 if α = 1, 2,3and uα = u2 if α = 4,5.

The limit paths xε (⋅, w) are all same, irrespective of the chosen subsequence or of w , because

the solution to (3.15) is unique under the chosen control u iε (⋅) . This implies that the sequence

J h ( x, u1h ,ε , u2h ,ε , α , ω 41 , ω 51 ) converges to the ε -optimal cost J ( x, u1ε , u2ε , α , ω 41 , ω 51 ) . Now,

using the optimality of V h ( x, α ) , we have:

limV h ( x, α ) ≤ lim J h ( x, u1h,ε , u2h,ε , α , ω 41 , ω 51 ) = J ( x, u1ε , u2ε , α , ω 41 , ω 51 ) ≤ V ( x, α ) + ε , (3.15.8)


h→0 h→0

Inequalities (3.15.7) and (3.15.8) imply that V h ( x, α ) → V ( x, α ).


REFERENCES

Akella, R. and Kumar, P.R., 1986. Optimal Control of Production Rate in a Failure-Prone
Manufacturing System. IEEE Transactions on Automatic Control AC 31, pp. 116-
12.
Boukas, E.K. and Haurie, A., 1990. Manufacturing Flow Control and Preventive
Maintenance: A Stochastic Control Aproach. IEEE Transactions on Automatic
Control 33, pp. 1024-1031.
Charlot, E., Kenné, J.P. and Nadeau, S., 2006. Optimal Production, Maintenance and
Lockout/tagout Control Policies in Manufacturing Systems. Int. J. Production
Economics 107, pp. 435-450.
Coit, D.W., 2001. Cold-standby Redundancy Optimization for Nonrepairable Systems. IIE
Trans, pp.471–478.
Ebeling, C.E., 1997. An introduction to reliability and maintainability engineering. New
York: McGraw-Hill Companies, Inc, pp.308.
Gershwin, S.B., 2002. Manufacturing Systems Engineering. Prentice-Hall, Englewood Cliffs,
NJ, pp.267.
Gharbi, A. and Kenné, J.P., 2003. Optimal Production Control Problem in Stochastic
Multiple-Product Multiple-Machine Manufacturing Systems. IIE Transactions 35,
pp. 941-952.
Gwo-Liang, L., Yen Hung, C. and Shey-Huei, S., 2009. Optimal Economic Production
Quantity Policy for Imperfect Process with Imperfect Repair and Maintenance.
European Journal of Operational Research, vol. 195, pp. 348-357.
Harichaux, P. and Libert, J., 2003. Ergonomie et Prévention des Risques Professionnels,
L'Environnement Physique du Travail et ses Contraintes, vol. 1, Chiron.
Kenné, J.P, Boukas, E.K. and Gharbi, A. 2003, Control of Production and Corrective
Maintenance Rates in a Multiple-Machine, Multiple-product Manufacturing System,
Mathematical and Computer Modelling 38, pp. 351-365.
Kimemia, J. and Gershwin, S.B., 1983. An Algorithm for the Computer Control of
Production in Flexible Manufacturing Systems. IEEE Transactions on Automatic
Control 15, pp.353-362.
Kushner, H.J. and Dupuis, P.G., 1992. Numerical Methods for Stochastic Control Problems
in Continuous Time. Springer, New York.
Kumar, S., and Kumar, D. and Mehta, N.P. 1996. Behavioural Analysis of a Shell
Gasification and Carbon Recovery Process in a Urea Fertilizer Plant. Microelectron
Reliab, pp. 671–673.
Nollet, J., Kelada, J., O.Diorio, M., Handfield, H., Deschamps, I. and R.Duguay, C., 1994. La
Gestion des Opérations et de la Production, Une Approche Systématique. 2e Édition,
Gaëtan Morin, pp. 163
60

Older, G.J and Suri, R., 1980. Time Optimal of Part-Routing in a Manufacturing System with
Failure Prone Machines. In: Proceedings of the 19th IEEE Conference on Decision
Control, vol. 5, pp. 722-727.
Pandey, D., Jacob, M. and Yadav, J. 1996. Reliability Analysis of a Power Loom Plant with
Cold-standby for its Strategic Unit. Microelectron Reliab, pp. 115–119.
Sinaki, G., 1994. Ultra-reliable Fault-tolerant Inertial Reference Unit for Spacecraft. In:
Proceedings of the annual rocky mountain guidance and control conference. Univelt
Inc., San Diego CA, pp. 239–48.
Subramanian, R., and Anantharaman V. 1995. Probabilistic Analysis of a Three Unit Cold
Standby Redundant System with Repair. Microelectron Reliab, pp.1001–1008.
Yan, H. and Zhang, Q., 1997. A Numerical Method in Optimal Production and Setup
Scheduling to Stochastic Manufacturing Systems. IEEE Transactions on Automatic
Control 42, pp. 1452-1455.
Zhang, Y.L., Wang, G.J. and Ji, Z.C. 2006. Replacement Problems for a Cold Standby
Repairable System, International Journal of Systems Science, pp.17–25.
Zhang, Y.L. and Wang, G.J. 2006. A Bivariate Optimal Repair-replacement Model Using
Geometric Process for Cold Standby Repairable System, Eng. Optim, pp.609–619.
Zhang, Y.L. and Wang, G.J. 2007. A Deteriorating Cold Standby Repairable System with
Priority in Use, European Journal of Operational Research, pp.278–295.
61

CHAPITRE 4

ARTICLE 2: LOCKOUT/TAGOUT AND OPTIMAL PRODUCTION


CONTROL POLICIES IN FAILURE-PRONE NON-HOMOGENOUS
TRANSFER LINES WITH PASSIVE REDUNDANCY

Behnam Emami-Mehrgani1, Jean-Pierre Kenné1, Sylvie Nadeau1


1
Départment of Mechanical Engineering, École de technologie supérieure, Montreal
(Quebec) H3C 1K3, Canada
This chapter has been submitted as an article for publication in International Journal of
Production Research. Accepted on January 26, 2012
Acceptance Confirmation Ms. Ref. No. : TPRS-2012-IJPR-0110.

Résumé

Ce présent article analyse une ligne de production constituée de trois machines produisant un
type de pièce (deux sous forme redondance passive et une troisième en série avec les
précédentes). Les machines sont sujettes à une dégradation, à des pannes, à des réparations
aléatoires. Le problème de contrôle considéré est soumis à des contraintes non-négatives des
inventaires en cours (work-in-process). La capacité du système est décrite par une chaîne de
Markov homogène à états finis. Les variables de décision sont les taux de production des
machines. Les variables de décision influencent le niveau des inventaires intermédiaires, des
inventaires de produits finis et la capacité du système manufacturier. L’objectif de cette
contribution est divisé en deux (2) volets : 1) Trouver les variables de décision permettant de
réduire les coûts totaux production, comprenant les coûts d’inventaire et de pénurie sur un
horizon infini de planification. 2) Intégrer la troisième machine sous forme redondance
passive afin de libérer un espace-temps essentiel pour minimiser les possibilités de
contournement des dispositifs de protection ou d’escamotage des procédures de
cadenassage/décadenassage (C/D). Afin d’avoir un modèle plus réaliste pour les industries,
les conditions d’optimum proposées sont développées en utilisant une approche combinée,
qui est présentée sur la base d'une combinaison de formalisme analytique, la simulation, le
62

plan d'expérience et la méthodologie de surface de réponse. L'utilité de l'approche proposée


est illustrée par un exemple numérique et une analyse de sensibilité.

Abstract

This paper considers a production problem for a transfer line subject to random failures and
repairs, and differs from other studies on transfer lines. It considers a manufacturing system
consisting of three machines (two machines with passive redundancy, and one in series with
the previous ones) producing one part type. The control problem is subject to non-negative
constraints on work-in-process (WIP). The decision variables are the production rates of two
main machines and a standby machine, and influence the WIP levels, the inventory levels
and the system’s capacity, which is assumed to be described by a finite-state Markov chain.
The objective of this paper is to minimize WIP and finished goods inventory costs; it also
aims to respect the essential space-time during intervention on machine down, in order to
minimize the possibility of the circumvention of protection devices or of the retraction of
lockout/tagout procedures through a passive redundancy system. This paper therefore verifies
the effect of passive redundancy on optimal stock levels. Given that an analytical or even a
numerical solution of the problem is very difficult to find, and that we want to have a more
realistic model for industries, we present a combined approach, which is presented based on a
combination of analytical formalism, simulation modeling, design of experiments and
response surface methodology to optimize a transfer line with passive redundancy, producing
one part type. The usefulness of the proposed approach is illustrated through a numerical
example and a sensitivity analysis.

Keywords: Production control; Lockout/tagout; Passive redundancy; Simulation,


Experimental design; Response surface methodology.

4.1 Introduction

To meet customer demand more and more demanding, and in the face of intense competition,
companies must be able to produce good quality products at lower cost. This requirement
63

necessarily makes performance a major concern for all managers. To achieve this goal, we
must ensure equipment availability, both functionally and in terms of optimal use. Production
systems optimization is a topic that is of interest to researchers as well as to industry. To
increase the availability of machines at the functional level, several maintenance strategies
have been developed (Kenne et al., 2003; Kenne and Gharbi, 1999; Rezg et al., 2004). The
problem of optimally controlling the production rates of a manufacturing system has been
widely discussed in the scientific literature. For two decades, significant effort has been
devoted to optimizing production systems to meet their complexity, competition and
challenges of globalization. However, despite the effectiveness of the techniques developed,
other challenges, including control and risk management of adverse events during
maintenance, still lie ahead.

Studies conducted in Quebec by the Institut de recherche Robert-Sauvé en santé et sécurité


du travail (IRSST), (Chinniah and Champoux, 2008) showed that in 2005, dangerous
machinery led to the deaths of about 20 workers in Quebec, and that there were 13,000
accidents linked to them. These accidents also caused $70 million in damages settled by the
Commission de la santé et de la sécurité du travail (CSST). Another study on the safety of
lockout/tagout of Sawmill equipment in Quebec conducted by the Institut de recherche
Robert-Sauvé en santé et sécurité du travail (IRSST), (Giraud et al., 2008) led the
occupational health and safety (OHS) community to recognize that nearly a quarter of all
accidents occur during interventions by workers on machines that are down. Manufacturing
systems operate in a stochastic environment because machines are subject to random
breakdowns and repairs, and in addition, demand for finished goods may vary. It is possible
to predict and control certain events while others occur randomly, and are beyond the control
of manufacturing systems (Gershwin, 2002). Production systems’ dynamics degrade with the
number of breakdowns and repairs, and with a rise in the number of breakdowns and repairs
comes a reduction in the availability of machines, as well as an increase in the number of
occupational hazards associated with maintenance activities; further, they can lead to serious
machine-related accidents, which are costly in terms of human life, sick leave days and
general financial costs. An important question arises: How can an optimal production policy
64

be maintained in an uncertain environment, while increasing the safety of maintenance


workers? The preferred method used to overcome these problems is the lockout/tagout. It
consists in locking a machine with a padlock in order to discharge all sources of residual
energy (hydraulic, electrical, etc.) in order to avoid premature starting of equipment
throughout a maintenance intervention.

Many managers and workers wrongly believe it takes too long to plan and carry out
lockout/tagout, thinking the accompanying downtime reduces productivity or performance.
Consequently, lockout/tagout is often deficient because idle production time is seen as a
hurdle to planned production rates. In this view, Charlot et al. (2006) considered an analytical
model combining lockout/tagout, production and corrective maintenance policies for a single
machine producing one part type. This work showed that lockout/tagout time can better be
controlled with the proper scheduling in production plan control. Another work integrating
lockout/tagout into operational risk in production control is proposed by Emami-Mehrgani et
al. (2011). They considered a manufacturing system with passive redundancy consisting of
two non-identical machines. Their work demonstrates clearly that passive redundancy
optimizes production and maintenance costs while enhancing occupational safety. Even
greater benefits occur if effective lockout/tagout maintenance planning occurs in concert with
production control.

Many authors have contributed to solving manufacturing systems production planning


problems. Based on Rishel’s formulation (Rishel, 1975), Older and Suri (1980) devised
manufacturing systems having unreliable machine control problems. In their model,
breakdowns and repairs are described by a homogeneous Markov process. The main
difficulty with this approach is its lack of efficient methods for solving the optimization
problem characterized by Hamilton-Jacobi-Bellman (HJB) equations. Akella and Kumar
(1986) analytically solved a one machine, one part-type problem. In this view, Lou and
Zhang (1994) extended the problem in Akella and Kumar (1986) to a flow control problem
for a tandem production system consisting of two unreliable machines, and conducted a
rigorous study of the dynamic properties of the system. Similarly, Presman et al. (1995)
65

considered a production planning problem in an N-machine flow-shop system subject to


machine breakdowns and repairs and to non-negativity constraints on work-in-process. Based
on the formulation presented in Presman et al. (1995), Hajji et al. (2009) studied production
and change-over control production for a buffered flow-shop producing several types of
parts. Hajji et al. (2009) have developed dynamic programming equations in terms of
problem directional derivative (DPEDD) and adopted a numerical approach to solve them.
The purpose of this paper is to control the production rate of the machines in a transfer line
with passive redundancy, which is one of the most important structures in reliability
engineering, and has been widely used in manufacturing systems. The main contributions of
this work are divided into two parts: reduce production costs and provide free space-time to
minimize the possibility of circumvention of protection devices or retraction of
lockout/tagout procedures for machines under repair. These goals are reached for a transfer
line through a passive redundancy system.

In this paper, by making use of the fact that the value function is the unique solution for the
associated Hamilton-Jacobi-Bellman (HJB) equations, in terms of directional derivatives
(DD), the structure of the solution under appropriate conditions is obtained. Given that an
analytical or even a numerical solution of the problem is very difficult to find and that we
want to have a more realistic model for industries, we present in this paper an alternative
procedure based on the combination of analytical control approach and the experimental
design method based on simulation experiments to find an approximation of the optimal
control policy. A simulation based experimental design approach is combined with the
control theory to develop a systematic control approach as in Gharbi and Kenne (2003) in the
case of production line with passive redundancy. The proposed control approach consists of
estimating the relationship between the incurred cost and the parameters of the control policy
considered in this paper as control factors. The hedging point policy, parameterized by these
factors, is used to conduct simulation experiments. For each configuration of input factors
values, the simulations model is used to determine the related output (incurred cost). An
input-output data set is then generated by simulation model. The significant effects of input
factors are determined by experimental design, and a response surface methodology is used
66

to obtain the relationship between the input and the output factors in order to estimate the
cost function. The best values for control factors are determined through this relationship.

This article is organised as follows. Assumptions and notations are defined in the next
section. In section 4.3, we provide the problem statement. Section 4.4 provides a numerical
example and a sensitivity analysis, and the related production policy is presented. In section
4.5, we present the control approach, the experimental design and response surface
methodology. Finally, we conclude the paper in section 4.6.

4.2 Assumptions and notations

This paper incorporates the following assumptions and notations:

4.2.1 Assumptions

1. Corrective maintenance is carried out with lockout/tagout.

2. The main machine is more robust than the standby machine.

3. The main machine and the standby machine produce the same type of parts for the work-
in-process (WIP).

4. The main machine returns to production immediately after each repair (corrective
maintenance with lockout/tagout) and the standby machine stands idle.

Assumption 4 is a classical assumption with a passive redundancy system, and is due to the
nature of a passive redundancy system.

4.2.2 Notations

The following notations are used in the rest of this article:

x1 (⋅) : inventory level of work-in-process

x2 (⋅) : inventory/backlog level of finished product

c1+ : holding cost incurred on buffer


67

c2+ : holding cost incurred on finished product

c2− : backlog cost incurred on finished product

cα : cost incurred for the operation of the machine under repair at mode α

cr1 : corrective maintenance cost of main machine M 1

cr2 : corrective maintenance cost of main machine M 2

crs : corrective maintenance cost of standby machine M s

ctagout : lockout/tagout cost

ur1 : corrective maintenance rate with lockout/tagout of main machine M 1

ur2 : corrective maintenance rate with lockout/tagout of main machine M 2

urs : corrective maintenance rate with lockout/tagout of standby machine M s

g (⋅) : instantaneous cost

J (⋅) : total cost

v (⋅) : value function

ρ: discount rate

d: demand rate

ui (⋅) : production rate of machine i ( i =1, 2, s )

uimax (⋅) : machine’s i ( i =1, 2, s ) maximal production rate

1,2
q12
: main machines M1 and M 2 failure rate

q12s : standby machine M s failure rate


68

4.2 Problem statement

In this paper, we consider a flow control problem for a tandem production system with
passive redundancy consisting of three unreliable machines. Two machines are in series
( M 1 and M 2 ) and one machine is in standby with the main machine M 1 , namely M S (standby
machine). Recall that in this paper, by using the fact that the value function is the unique
solution for the associated Hamilton-Jacobi-Bellman (HJB) equations, in terms of directional
derivatives (DD), the structure of the solution under appropriate conditions is obtained. Since
either an analytical, a numerical solution of this problem or even an explicit functional
relationship between the independent variables of the model (stock level) and performance
criteria (cost incurred) is not usually available, an alternative procedure based on the
combination of analytical control approach and the experimental design method based on
simulation experiments is presented in this paper. A parameterized near-optimal control
policy is used in the proposed control approach as input for the simulation model. In order to
propose an approach which could be easily applied to control manufacturing systems for a
tandem production system with passive redundancy. The manufacturing system is consisting
of three unreliable machines (two machines with passive redundancy, and one in series with
the previous ones) at the operational level, the descriptive capacities of discrete event
simulations models are combined with analytical models, experimental design, and response
surface methodology. The system is shown in Figure 4.1. The main machines have two
states: ξ1,2 (t ) = 1 if the main machine M i ( i = 1, 2) is operational and ξ1,2 (t ) = 2 if the main

machine M i (i = 1, 2) is under repair. The standby machine has three states: ξs (t ) = 1 if the

standby machine M s is operational, ξ s (t ) = 2 if the standby machine M s is under repair and

ξs (t ) = 3 if the standby machine M s is at time-off. Hence, the dynamics for a manufacturing


system consisting of three machines (two machines with passive redundancy and one
machine in series with the previous ones) is in a hybrid state comprised of a continuous state
and a discrete state ξ (t ) = (ξ (t )1 , ξ (t )2 , ξ (t )s ) as follows:
69

a) Continuous state: We denote the number of parts in the work-in-process (WIP) as x1 (t )

and the difference between cumulative production and demand as x2 (t ) . Note that the
control problem is subject to non-negative constraints on work-in-process (WIP), meaning
that x1 (t ) ≥ 0. The surplus x2 (t ) can be positive (i.e., inventory costs c2+ are thus charged)

or negative (i.e. backlog costs c2− are thus charged). We use ui (t ) with (i = 1, 2, s ) to denote

the input rate to M i and x1 (t ) to denote the number of parts in the buffer between

M i and M 2 (i = 1, s) .

Figure 4.1 Transfer line with passive redundancy producing one part type

The dynamics of the system can be written as follows:

x1 (t) = u1(t) − u2 (t) =u1(t)Ι1α + us (t)Ιαs −u2 (t), x1 (0) = x1 , (4.1)
with:
1 if α = 3,4,5,6,7,8, 1 if α = 1,2,
Ι1α =  and Ιαs = 
0 otherwise, 0 otherwise.

Note that Ι1α and Ιαs present the state of main machines and standby machine in each mode.

x2 (t) =u2 (t) −u3 (t) = u2 (t) −u3 (t), x2 (0) = x2 , (4.2)
with:
u3 (t):= d.
70

In matrix notation, the system of equation (4.1)-(4.2) becomes:

x(t) = Au(
 t), x(0) = x, (4.3)

1 −1 0  α α
where A=   , u(t) = (u1(t), u2(t), u3(t)) = (u1(t)Ι1 + us (t)Ιs , u2(t), u3(t)) and x(t) = (x1(t), x2(t)),
0 1 −1
with:
1 if α = 3,4,5,6,7,8 1 if α =1,2,
Ια =  , Ια
s =  and u (t):= d.
1 0 otherwise 0 otherwise,
3

b) Discrete state: Consistent with assumptions set out in section 4.2.1, the operational mode
of the whole system can be described by a random vector ξ (t ) = (ξ1 (t ), ξ 2 (t ), ξ s (t )) taking

values in Β = {1, 2,3, 4,5, 6, 7,8} . Without loss of generality, for the three-machine flow-

shop case with passive redundancy, ξ (t ) can be expressed as follows:

1 M1 is under repair, M 2 is operational and M S is operational;


2 M1 is under repair, M 2 is under repair and M S is operational;

3 M1 is operational, M 2 is operational and M S is under repair;

4 M1 is operational, M 2 is under repair and M S is under repair;
ξ (t ) = 
5 M1 is under repair, M 2 is operational and M S is under repair;
6 M1 is under repair, M 2 is under repair and M S is under repair;

7 M1 is operational, M 2 is operational and M S is at time-off;
8 M1 is operational, M 2 is under repair and M S is at time-off.

Figure 4.2 displays the modes of the system associated to the process ξ (t ) .
71

Figure 4.2 State transition diagram

The transition rate matrix of the stochastic processes ξ (t ) is denoted by Q such that

Q = {qαβ } , with qαβ > 0 if α ≠ β and qαα = − β ≠α qαβ , where α , β ∈ Β.

The transition probabilities associated to qαβ are expressed as:

qαβ (⋅)δt+0(δt) if α≠β,



p ξ(t +δt) = β ξ(t) =α =  (4.4)
1+q (⋅)δt+0(δt) if α=β.
 αβ

The transition rate matrix Q is expressed as follows:

 q11 q12 0 0 q15 0 q17 0


q q22 0 0 0 q26 0 q28 
 21
0 0 q33 q34 q35 0 q37 0
 
0 0 q43 q44 0 q46 0 q48 
, (4.5)
 q51 0 q53 0 q55 q56 0 0
 
0 q62 0 q64 q65 q66 0 0
q 0 q73 0 0 0 q77 q78 
 71 
 0 q82 0 q84 0 0 q87 q88 
72

The set of admissible decisions at mode α (t ) and control policies (control variables) at mode
α (t ) :
((u1(⋅), u2 (⋅), us (⋅)) ∈R3, 
 
Γ(x,α) = 0 ≤ u1(⋅) ≤ u1max ,0 ≤ u2 (⋅) ≤ u2max , , (4.6)
0 ≤ u (⋅) ≤ umax 
 s s 

max max
In equation (4.6), u
1
is the main machine’s M 1 maximal production rate, u
2
is the main

machine’s M 2 maximal production rate and u smax is the standby machine’s M s maximal

production rate. Γ ( x, α ) denotes the set of all admissible controls with respect to x ∈ Λ and

α (0) = α . Let Λ = [ 0, ∞ ) × R ⊂ R m denote the state constraint domain.

The control problem consists in finding an admissible control law u(⋅) = (u1, u2 , us ) that

minimizes the cost function J (⋅) given by:

∞ −ρ t 
J(α, x,u) =E e g(α, x,⋅) dt x( 0) =x, ξ( 0) =α, (4.7)
0 

where ρ is the discount rate and g(x, α , ⋅) = c1+ x1+ + c2+ x2+ + c2− x2− + cα is the instantaneous cost,

c + , c − and cα , being the cost per unit to produce parts for inventory, backlog as well as
intervention cost on the machine, respectively.

x+ = max {0, x} , x− = max {− x,0}


and

c α = (( c r1 + ctagout )u r1 )Ind {α = 1} + (( c r1 + ctagout )u r1 + ( c r2 + ctagout )u r2 )Ind {α = 2}


+ (( c rs + ctagout )u rs )Ind {α = 3} + (( c rs + ctagout )u rs + ( c r2 + ctagout )u r2 )Ind {α = 4}
+ (( c rs + ctagout )u rs + ( c r1 + ctagout )u r1 )Ind {α = 5}
+ (( c rs + ctagout )u rs + ( c r1 + ctagout )u r1 + ( c r2 + ctagout )u r 2 )Ind {α = 6}
+ (( c r2 + ctagout )u r2 )Ind {α = 8}
73

With:
1 if Θ(⋅) is true 
Ind {Θ(⋅)} =  
0 otherwise 

For a given proposition Θ(⋅) . The cost of corrective maintenance with logout/tagout depends
on the duration of lockout/tagout and repair activities. This cost is described for main
machine M 1 by (cr1 + ctagout )ur1 , for the main machine M 2 by (cr2 + ctagout )ur2 and for the

standby machine M s by (crs + ctagout )urs .

Letν ( x,α ) denote the value function or minimum discounted cost for equations (4.7) as
expressed in the following equation:

ν ( x,α) = min J (α, x, u) , ∀α ∈Β (4.8)


u∈Γ( x,α)
In Appendix 4.A, we present the properties of the value function v(⋅) given by equation
(4.8). It is shown that the value function v(⋅) given by (4.8) should satisfy a set of partial
differential equations known as the Hamilton-Jacobi-Bellman (HJB) equations in terms of
directional derivatives (DD).

4.3 Numerical example and sensitivity analysis

Let us consider a transfer line with three non-identical machines (two machines with passive
redundancy and one machine in series with the previous ones). The system capacity is
described by an eight Markov process with states ξ (t ) ∈ Β = [1,2,3,4,5,6,7,8] .

The discrete dynamic programming equation (4.A.6) in Appendix 4.A gives the eight
equations which are presented in Appendix 4.B.

We use the following computational domain:

Gxh = {( x1 , x2 ) : 0 ≤ x1 ≤ 5; − 5 ≤ x2 ≤ 5} ,
The parameters for our case study appear in Table 4.1.
74

Table 4.1 Parameters of the numerical example

The results obtained for the control variables u1 , u2 and us of a production line with a passive
redundancy system are given in Figures (4.3)-(4.9) for illustration purposes.

(a) Production rate of M s (b) Production rate of M 2

Figure 4.3 Production rate of M s and M 2 at mode 1

(a) Production rate of M 1 (b) Production rate of M 2

Figure 4.4 Production rate of M 1 and M 2 at mode 3


75

a) Production rate of M 1 (b) Production rate of M 2

Figure 4.5 Production rate of M1 and M 2 at mode 7

Figure 4.3 shows that there is no need to produce the part with sufficient stock levels both in
the work-in-process (WIP), described by x1 and the stock of finished products, described by
x2. In mode 1, the production rate of the main machine ( M1 ) is described by u1 = 0 . For
small stock levels, the policy obtained properly defines the region in the domain (x1,x2)
where a maximal production rate is optimal. In Figure 4.4, we have a similar trend in the
optimal production rate as in mode 1. In mode 3, the production rate of the standby machine
(M s ) is described by us = 0 . Figure 4.5 illustrates the same policy as in modes 1 and 3, just
that at mode 7, the standby machine is at time-off, meaning that the production rate of
standby machine us = 0 .
76

Production rate of M s

Figure 4.6 Production rate of M s at mode 2

Production rate of M 1

Figure 4.7 Production rate of M1 at mode 4


77

Production rate of M 1

Figure 4.8 Production rate of M 1 at mode 8

Figure 4.6 illustrates that thanks to a standby machine, we can produce the part for the work-
in-process (WIP), as described by x1. In mode 2, the production rates of the main machines
M 1 and M 2 are described by u1 = u2 = 0 respectively. Figure 4.7 shows that the main

machine M 1 produces parts for the work-in-process (WIP) described by x1. In mode 4, the

production rates of main machine M 2 and of standby machine M s are described by

u2 = us = 0 respectively. We have the same policy in Figure 4.8 as in modes 2 and 4, except
that at mode 8, the standby machine is at time-off, meaning that the standby machine’s
production rate is us = 0 . The production rate of the main machine M 2 at mode 8 is

described by u2 = 0 .
78

Production rate of M 2

Figure 4.9 Production rate of M 2 at mode 5

Figure 4.9 shows that the main machine M 2 produces the finished products described by x2.

In this mode, the production rates of main machine M 1 and of standby machine M s are

described by u1 = us = 0 .

These results also illustrate that the main machines and standby machine are unavailable at
mode 6, which means that u1 = u2 = us = 0 .

The results illustrated in Figures (4.3)-(4.9) allowed us to determine our policy as follows:

 uimax if x1 < Z1 & x2 < Z 2 


  
ui ( x1 , x2 ,3) =  d if x1 = Z1 & x2 = Z 2 
  
  0 otherwise 
 , (4.9)
  u max
2
if x 2 < Z 3 & x1 > 0  
 u ( x , x ,3) =  d = >

 2 1 2  if x2 Z 3 & x 1 0 
  
 0 otherwise 
With i = 1, S
79

Note that the aim of our policy is to find the optimal value of production rates ui with i = 1, S

which are dependent on two factors, Z1 and Z2 , and the optimal value of production rate u2

which is dependent on one factor such Z3 , as illustrated in Figure 4.9.

The next sections are aimed at developing a systematic approach for determining the optimal
values of Z1 , Z2 , Z3 .

4.4 Control approach, experimental design and response surface methodology

In order to propose an approach which could be easily applied to control manufacturing


systems at the operational level, the descriptive capacities of discrete event simulations
models are combined with analytical models, experimental design, and response surface
methodology. Many studies have been covered in the research literature on these subjects.
For more details, we refer the reader to Gharbi and Kenné (2000). The structure of the
proposed control approach is presented in Figure 4.10.

FMS Control Analytical Numerical


problem model methods

Z1* , Z 2* , Z3* Z1 , Z 2 , Z3

Response Experimental Simulation


Figure 3.10 Diagram
surface of control approach
design model

Figure 4.10 Diagram of control approach

1. The manufacturing system’s control problem statement, as shown in section 4.3, consists
of a production problem presentation for a transfer line with passive redundancy. This
problem is presented through a stochastic optimal control model based on control theory.
80

The aim of this step is to find the control variables (u1(⋅), u2 (⋅), us (⋅)) , called the
production rates. The control variables allow an improvement of the incurred cost.
2. The optimality conditions, described by the HJBDD equations, are obtained from the
problem statement of the first step. This step shows that the value function, representing
the incurred cost, is the solution of the HJBDD equations, and our control policy
(production rates) is near-optimal.
3. In this step, we use numerical methods to solve the optimality equations of the problem
because it is not possible to solve them analytically.
4. The control factors Z1 , Z2 , Z3 for the control production rates describe the numerical control
policy obtained.
5. The simulation model uses the near-optimal control policy defined in the previous step as
the input factor for conducting experiments in order to evaluate the transfer line’s with
passive redundancy performances. Therefore, the cost incurred is obtained for the given
values of the control factors thanks to the simulation model which will be presented in the
next section.
6. The experimental design approach defines how control factors can be varied in order to
identify the effects of the main factors and their interactions on the cost. These variations
must be evaluated through a minimal set of simulation experiments.
7. In this step, we use a response surface methodology to obtain the relationship between the
significant main factors and the incurred cost as well as the relationship between the main
factors’ interactions. Thereafter, the optimized model obtained in order to determine the
main factors’ best values are called Z1* , Z2* , Z3* for the production.
8. Using the proposed control approach gives the production rates described by equation
(4.9), for the best values of factors Z1 , Z 2 , Z 3 meaning Z1* , Z2* , Z3* .

The performance-estimation tool chosen for this study is a discrete simulation model, and so
we used the Arena software, which uses the SIMAN language; we refer the reader to Rossetti
(2010).
81

The SIMAN portion of the software is composed of various networks describing specific
tasks (failure and repair events, threshold production crossing of inventory variables, etc.).
The simulation model is presented in Figure 4.11.

Figure 4.11 Diagram of the simulation model


1. The INITIALIZATION block initializes the problem variables (current surplus, production
rates, incurred cost, etc.).

2. The DEMAND ARRIVAL block performs the arrival of a demand for each d unit of time.
A verification test is then performed on the product’s inventory level, and the inventory or
the backorder is updated.
82

3. The CONTROL POLICY block is defined in equation (4.9). The feedback control policy
is defined by the output of the SIGNAL block, which is used to permanently verify the
variation in the stock level x (t ) in order to specify the best action to carry out.

4. The STARVATION of the machines is implemented with the use of observation networks.
Whenever the in-process buffer becomes empty, a signal is sent. Another signal is sent
when the material becomes available for operation

5. The FAILURES AND REPAIRS block performs two functions: it defines the time-to-
failure of the machine as well as the time-to-repair of the machine.

6. The PARTS PRODUCTION block performs the production of finished products according
to the policy defined by the CONTROL POLICY.

7. The UPDATE INVENTORY block is used once the time step is chosen. For more details
we refer the reader to Pritsker (1999).

8. The UPDATE INCURRED COST block calculates the incurred cost according to the
different variable levels and unit costs c+ and c − .

The simulation ends when the current simulation time t reaches the defined simulation period
T. We thus ran offline simulations to determine the time necessary for the manufacturing
system to reach its steady state. We found the simulation time for our manufacturing system
at nearly 20000 units of time, and this duration is used for all the simulations.

The simulation parameters used in this paper are the same as in Table 4.1. The input-output
data is generated by the simulation model from the variation of independent factors
Z1 , Z 2 and Z3 . Now, we present a procedure to vary these factors simultaneously. Such a

procedure uses the experimental design approach. Hence, for three factor problems, as
illustrated in the previous sections, we selected a 33 response surface design since we have
three independent variables at three levels. This design leads to the completion of 27
experimental trials. The levels of the independent variables Z1, Z2 , Z3 range from a low of 5 to
a high of 25. In this paper, we chose three replications, and thus have 108 (27*4) simulation
83

runs. We refer the reader to Montgomery (2005) for more details. We considered all possible
combinations of different levels of independent variables by response surface design. The
objective of this design is to understand the effects of independent variables on performance
measures; in our case, the production average cost. From the ANOVA table, the independent
variables Z1 , Z 2 , Z3 and interaction effect as well as the quadratic effect are significant for the
dependant variable at 0.05 level of significance. The R2 value of 0.88, meaning 88% of the
total variability, is explained by the model.

The average cost function is given by:


n n n
Average cost = β0 +  βi Zi +  βij Zi Z j , (4.10)
i =1 i =1 j =1
j ≥i

The estimation of the regression coefficients is performed and ten coefficients achieved in
Table 4.2.

Table 4.2 Polynomial coefficients

The minimum average cost function is located at Z1* = 11.4, Z 2* = 14.5, Z3* = 13 ,where

Z1* , Z 2* and Z 3* represent the optimal values of independent variables Z1 , Z 2 and Z3 .

Figures (4.12)-(4.14), illustrate the contour plots of the average cost function or response
surface.
84

Figure 4.12 Contour plot of the response surface

Figure 4.13 Contour plot of the response surface


85

Figure 4.14 Contour plot of the response surface

These values determine the extension of the hedging point policy for the manufacturing
system considered, where the average cost is minimised and this control policy is the best
approximation of the optimal control. We observe that the cost function is not very sensitive
to small variations of finished goods stock levels. The work-in-process (WIP) stock levels
appear to be less sensitive to increases or decreases in WIP stock levels’ small variations.
Because we can produce parts for work-in-process (WIP) at all times thanks to passive
redundancy, we can therefore respond to demand permanently.

To illustrate the effect of the variation inventory and backlog costs on the design parameters,
a sensitivity analysis was conducted in Table 4.3, with c1+ = 1, cr1 = 150,

cr2 = 200, crs = 250 and ctagout = 50.


86

Table 4.3 Variations of optimal design factors based on inventory and backlog costs

The first section of Table 4.3 presents the variations in stock levels based on inventory costs,
while the second section shows the variations in stock levels based on backlog costs.

Figures 4.15 and 4.16 plot variations in stock levels based on inventory costs and backlog
costs, which are presented in Table 4.3. In these figures, homogenous variation occurs with
passive redundancy, because it produces parts for work-in-process (WIP) at all times and
makes possible to respond to demand permanently. Therefore, we observe that variations in
inventory and backlog cost do not influence the production thresholds of work-in-process
(WIP) and finished goods.
87

Figure 4.15 Production threshold /Inventory cost

Figure 4.16 Production threshold/Backlog cost

Production a transfer line’s optimal costs was determined using the analytical model
presented in this paper. The numerical approach, the experimental design and the response
88

surface methodology consecutively show that the resulting policy is optimal and enhances
machine availability. Without loss of generality of this proposal, this model is based on
certain assumptions relating to a transfer line consisting of three machines which are not
identical, and operate with passive redundancy. We observed that the passive redundancy
case allows us to better optimize the production cost of transfer lines while guaranteeing
occupational safety. The integration of the second machine as the passive redundancy allows
demand to be met permanently. Furthermore, this integration will release the intervention
time needed for the machine which is down. It also minimises the possibility of
circumvention of device protection or retraction of lockout-tagout procedures.

4.5 Conclusion

This paper confirms that it is possible to integrate a passive redundancy system in a


production line in order to: 1) increase the productivity of workers and material resources, 2)
better optimize production costs while guaranteeing the safety of workers. In this paper, the
control policy has an extension of hedging point structure. Based on the numerical solution
obtained, a parameterized near-optimal control policy was derived. Such a policy depends on
stock threshold levels. An experimental design was used to determine the effects of the
independent variables on the average cost over the production horizon. We combined an
analytical, simulation and statistical method to provide the average cost estimation related to
the control problem. Average cost’s estimation permits to know the best values of the control
parameters. Finally, passive redundancy system improves production costs and worker
safety. The former is obtained by meeting the demand whereas the latter is fulfilled by
releasing essential space-time for the machines that are under repair.

Acknowledgments

We gratefully acknowledge the financial support for this work from the following agencies:
Natural Sciences and Engineering Research Council of Canada (NSERC); Réseau de
Recherche en Santé et en Sécurité du Travail du Québec (RRSSTQ); Fonds de la Recherche
en Santé – Québec (FRSQ); Fonds Québécois de la Recherche sur la Société et la Culture
89

(FQRSC); Fonds Québécois de la Recherche sur la Nature et les Technologies (FQRNT);


Institut de Recherche Robert-Sauvé en Santé et en Sécurité du Travail (IRSST) and Équipe
de Recherche en Sécurité du Travail (EREST).
90

APPENDIX 4.A OPTIMAL CONDITIONS AND NUMERICAL APPROACH

The properties of the value functions and the Hamilton-Jacobi-Bellman equation (HJB) in
terms of directional derivative (DD) for inner and boundary points are presented in this
section. These equations describe the optimality conditions for a transfer line with passive
redundancy producing one part type. Hence, we first present the notion of these derivatives
and some related properties of convex functions.

A function f ( x), x ∈ Rm , is said to have a directional derivative f p' ( x ) along direction

p ∈ Rm , if there exists
f ( x + δ p) − f ( x)
lim = f p' ( x ).
x→0 δ
If a function f ( x) is differentiable at x , then f p' ( x ) exists for every p and

f p' ( x ) = ∇f ( x ), p

Where ∇f ( x ) is the gradient of f ( x ) and .,. is the scalar product. Furthermore, a

continuous convex function defined on a convex domain  is differentiable almost

everywhere, and has a directional derivative along any direction at any inner point of 
and along any admissible direction (i.e., a direction p such that x + δ p ∈  for some δ > 0 )

at any boundary point of  (for more details, see Sethi and Zhang, 1994).

Note that { Au : (u1 , u2 , u s ) ∈ Γ ( x, α )} is the set of admissible directions at x .

Regarding the optimality principle, let us write the set of partial differential equations known
as the Hamilton-Jacobi-Bellman (HJB) equations in terms of directional derivatives (DD) as
follows:

 ' 
ρν ( x,α ) = min vAu ( x, α ) + g (α , x, ⋅) +  qαβ ν ( x,β ) −ν ( x,α )   , (4.A.1)
u∈Γ( x,α )  α ≠β 
∀α , β ∈ Β
91

where:

v A' u ( x, α ) = (u1 − u2 )vx1 ( x, α ) + (u2 − u3 ) × vx2 ( x, α ).

with:
1 if α = 3,4,5,6,7,8, 1 if α = 1,2,
u1 = u1Ι1α + usΙαs , Ι1α =  , Ιαs =  , u2 = u2 and u3 = u3 := d.
0 otherwise, 0 otherwise,

The choice of standby machine and the main machine characteristics must be in such a way
that respects the feasibility of system.

The system is considered feasible if:


max ≥ d ,
π ui (4.A.2)
Where the limitation probabilities can be ascertained from the following equation for a
system conforming to a Markov process:

π (⋅)Q(⋅) = 0,

π = 1
with:

π (⋅) : Limiting probabilities

Q (⋅) : Transition matrix rates

Hence, we have π (⋅) = (π1 , π 2 , π 3 , π 4 , π 5 , π 6 , π 7 , π 8 ) representing the vector of limiting


probabilities from modes 1 to 8.

Let δΛ denote the boundary of Λ . If there exists xi = 0 (i = 1, 2), then x ∈ δΛ . Let the

restriction of ν ( x,α ) on some j-dimensional face, 0 < j < m , of δΛ be differentiable at an

inner point x0 of this face. Hence, there is a vector  v( x , α )


∇ such that
0

 v( x , α ), p for any admissible direction at x .


v 'p ( x0 , α ) = ∇ 0 0

Now, we can write the boundary condition on ν ( .,.) from the continuity of the value function

by:
92

  
min  ∇v( x0 , α ), Au + g (α , x0 , ⋅) +  qαβ ν ( x0 ,β ) −ν ( x0 ,α )  
u∈Γ( x0 , α )  α ≠β 
  
= min  ∇ v ( x0 , α ), Au + g (α , x0 , ⋅) +  qαβ ν ( x0 ,β ) −ν ( x0 ,α )   , (4.A.3)
u∈Γ(α )  α ≠β 
∀α , β ∈ Β

We refer the reader to Lou and Zhang (1994) for the interpretation of the condition (4.A.3).
The optimal control policy (u1* , u2* , us* ) denotes a minimizer over Γ( x,α ) of the right-hand
side of equation (4.A.1). This policy is consistent with the value function obtained in
Equation (4.8). The optimal control policy therefore rests in solving Equation (4.A.1).
Obtaining an analytical solution of equation (4.A.1) is roughly impossible. The numerical
solution of the HJB equation (4.A.1) in terms of directional derivatives (DD) is a challenge
considered insurmountable in the scientific literature.

Now, we use numerical methods to solve the optimality conditions presented in this section.
This method is based on Kushner’s approach (Kushner and Dupuis, 1992). The basic idea
behind this approach consists in using an approximation scheme for the directional derivative
of the value function v ( x, α ). Let h1 and h2 denote the length of the finite difference interval

of the variables x1 and x2 . Hence, using h1 and h2 , v ( x , α ) is approximated by vh ( x,α ) , and

vx and vx are approximated by:


1 2

1 h 
 h (ν ( x1, x1 + h1,α ) −ν ( x1,α ) ) ×(u1 −u2 ) if ( u1 −u2 ) ≥ 0,
h

 1 
1 h 
 h (ν ( x1,α ) −ν ( x1, x1-h1,α ) ) ×(u1 −u2 )
h
'
otherwise, 
vx1 (⋅) = vu1−u2 (⋅) =  1 , (4.A.4)
 with: 
 
u = u Ια + u Ια , Ια = 1 if α = 3,4,5,6,7,8, , Ια = 1 if α =1,2, and u2 = u2 , 
 1 1 1 s s 1 0 otherwise,
s  
  0 otherwise, 
93

1 h 
 h (ν ( x1, x2 + h2 ,α ) −ν ( x1,α ) ) ×(u2 − u3 )
h
if (u2 − u3 ) ≥ 0, 
 2 
1 h 
vx2 (⋅) = vu'2 −u3 (⋅) =  (ν ( x1,α ) −ν ( x1, x2 -h2 ,α ) ) ×(u2 − u3 )
h
otherwise, , (4.A.5)
 2h 
 with: 
 
 u2 = u2 and u3 = u3 := d, 

We manipulated the approximation arrived at in equations (4.A.4) and (4.A.5) to rewrite the
HJBDD equations (4.A.1) as follows:

  
h
   qαβ (v (x, β)) + g(x,.) 
 β≠α 
−1
 (u1 −u2) (u2 −u3) (u1 −u2) h 
ν ( x,α) = min ρ + qαα +
h
+   (ν ( x1 +h1, x2,α) k1 +ν ( x1 -h1, x2,α) k1 ) ,
+ h −
(4.A.6)
u∈Γh( x,α)  h1 h2  h1
 
  (u2 −u3) h − 
 +
 h  (ν ( x1, x2 +h2,α) k+
2 +ν h
( x1, x2 −h2,α) k2 )
  2 

with:
1 if α = 3,4,5,6,7,8, 1 if α = 1,2,
u1 = u1Ι1α + usΙαs , Ι1α =  , Ιαs =  , u2 = u2 and u3 = u3 := d.
0 otherwise, 0 otherwise,

Where Γ h  x,α  is the discrete feasible control space and the other terms used in equation

(4.A.6) are defined as:

1 if (u1 − u2 ) ≥ 0, 1 if (u1 − u2 ) < 0,


K1+ =  and K1− = 
0 otherwise, 0 otherwise,

with:
1 if α = 3, 4,5, 6, 7,8, 1 if α = 1, 2,
u1 = u1Ι1α + us Ιαs , Ι1α =  , Ιαs =  and u2 = u2 .
0 otherwise, 0 otherwise,
94

1 if (u2 − u3 ) ≥ 0, 1 if (u2 − u3 ) < 0,


K 2+ =  and K 2− = 
0 otherwise, 0 otherwise,
with:
u2 = u2 and u3 = u3 := d .

In this paper, we use the policy improvement technique to derive an approximate optimization
problem solution. The algorithm of this technique can be found in Kushner and Dupuis (1992).
95

APPENDIX 4.B

The discrete dynamic programming equation (4.A.6) in Appendix 4.A gives the following
eight equations:

 us −u2 (νh( x +h, x ,1) k+ +νh( x -h, x ,1) k−) 


 h1 1 1 2 1 1 1 2 1 
−1 
 u −u2 u2 −d   u2 −d h 
νh( x,1) = min  ρ+ s + +q12 +q15 +q17  + (ν ( x1, x2 +h2,1) k2+ +νh( x1, x2 −h2,1) k2−), (4.B.1)
u∈Γh( x,1)  h1 h2   h2 
 h h h 
+g( x,1) +q12ν ( x,2) +q15ν ( x,5) +q17ν ( x,4) 
 

 us −u2 (νh ( x +h, x ,2) k+ +νh ( x - h, x ,2) k−)


 h 1 1 1
1
 2 1 1 2

 1 
−1
 u −u2 d   d 
νh( x,2) = min  ρ + s + +q21 +q26 +q28  + (νh ( x , x −h ,2) k2−) + g( x,2) +q21νh ( x,1) , (4.B.2)
u∈Γh( x,2)  h1 h2   h2
1 2 2


 
+q26νh( x,6) +q28νh( x,8) 
 

 u1 −u2 h h 
 h (ν ( x +h, x ,3) k1 +ν ( x -h, x ,3) k1 ) 
+ −
1 1 2 1 1 2

 1 
−1 
 u1 −u2 u2 −d   u −d 
νh( x,3) = min  ρ+ + +q34 +q35 +q37  + 2 (νh( x, x +h,3) k+ +νh( x, x −h,3) k−),(4.B.3)
u∈Γh( x,3)  h1 h2   h2 2 1
2

2 2 1 2 2

 
+g( x,3) +q34νh( x,4) +q35νh( x,5) +q37νh( x,7) 
 

 u −u (νh( x +h, x ,4) k+ +νh( x -h, x ,4) k−)


1 2 
 h 1
1
1
1 2 1 1 2

−1 
1

 u −u d  d 
νh( x,4) = min  ρ + 1 + +q43 +q46 +q48   (νh( x, x −h,4) k−)+g( x,4) +q43νh( x,3)
2
, (4.B.4)
u∈Γh( x,4 )  h1 h2  h2 2
1 2 2


 h h 
+q46ν ( x,6) +q48ν ( x,8) 
 
96

u2 (νh ( x - h , x ,5) k−) 


 h1 1 1 2 1

−1 
h  u2 u2 −d   u2 −d h ( h − 
ν ( x,5) = min  ρ + + +q51 +q53 +q56  + +
(ν x1 +h1, x2,5) k2 +ν ( x1, x2 −h2,5) k2 ), (4.B.5)
u∈Γh( x,5)  h1 h2   h2 
 h h h 
+g( x,5) +q51ν ( x,1) +q53ν ( x,3) +q56ν ( x,6) 
 

d h − 
 h (ν ( x1 , x2 − h2 ,6 ) k2 ) 
−1  2 
h  d   h 
ν ( x,6 ) = min  ρ + + q62 + q64 + q65   + g ( x,6 ) + q62ν ( x,2 )  , (4.B.6)
u∈Γh ( x ,6 )  h2   
 +q64ν h ( x,4 ) + q65ν h ( x,5 ) 
 
 

 u1 −u2 h + h − 
 h (ν ( x1 +h1, x2,7) k1 +ν ( x1 -h1, x2,7) k1 ) 
−1 1 
 u −d 
νh( x,7) = min  ρ+ 1 2 + 2 +q71 +q73 +q78  + 2 (νh( x1, x2 +h2,7) k+ +νh( x1, x2 −h2,7) k−)
u −u u −d
2 2 , (4.B.7)
u∈Γh( x,7)  h1 h2   h2 
 h h h 
+g( x,7) +q71ν ( x,1) +q73ν ( x,3) +q78ν ( x,8) 
 

 u1 −u2 (νh( x +h, x ,8) k+ +νh( x -h, x ,8) k−) 


 h 1
1
1 2 1 1
1
2

−1 1 
h  u1 −u2 d   d h 
ν ( x,8) = min  ρ+ + +q82 +q84 +q87  + (ν ( x, x −h,8) k−)+g( x,8)
 , (4.B.8)
h
u∈Γ ( x,8)  h h   h2
1 2
2
2


1 2
 
+q82νh( x,2) +q84νh( x,4) +q87νh( x,7) 
 
97

REFERENCES

Akella, R. and Kumar, P.R., 1986. Optimal Control of Production Rate in a Failure-Prone
Manufacturing System. IEEE Transactions on Automatic Control AC 31, pp. 116-
12.
Charlot, E., Kenné, J.P. and Nadeau, S., 2006. Optimal Production, Maintenance and
Lockout/tagout Control Policies in Manufacturing Systems. International Journal of
Production Economics 107, pp. 435-450.
Chinniah, Y., and Champoux, M., 2008. La sécurité des machines automatisées : Analyse des
risques et des moyens de protection sur une presse à injection de plastique. R-557.
Montréal: IRSST, pp. 93.
Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011. Lockout/tagout and operational risks
in the production control of manufacturing systems with passive redundancy.
International Journal of Production Economics 132, pp. 165-173.
Gershwin, S.B., 2002. Manufacturing Systems Engineering. Prentice-Hall, Englewood Cliffs,
NJ, pp. 267.
Gharbi, A., and Kenné, J.P. 2000. Production and preventive maintenance rates control for
manufacturing system: an experimental design approach. International Journal of
Production Economics, 65(3), pp. 275-287.
Giraud, L., Ledoux, É., Tanchoux, S and Pâques, J.J., 2008. La maintenance, État de la
connaissance et étude exploratoire. R-578. Montréal: IRSST, pp. 47.
Hajji, A., Gharbi, A., Kenné, J.P., and Hidehiko, Y. 2009. Production and changeover control
policies of a class of failure prone buffered flow-shops. Production Planning &
Control, Vol. 20, No. 8, pp. 785-800.
Kenne, J.P., E.K. Boukas and Gharbi, A. 2003. Control of production and corrective
maintenance rates in a multiple-machine, multiple-product manufacturing system.
Mathematical and Computer Modelling, Vol. 38, No. 3-4, pp. 351-365.
Kenne, J.P., Gharbi, A., 1999. Experimental design in production and maintenance control
problem of a single machine, single product manufacturing system. International
Journal of Production Research, Vol. 37, pp. 621-637.
Kushner, H.J. and Dupuis, P.G., 1992. Numerical Methods for Stochastic Control Problems
in Continuous Time. Springer, New York.
Lou, S.P. and Zhang, Q., 1994. Optimal Feedback Production Planning in a Stochastic Two-
Machine Flowshop. European Journal of Operational Research, 73, pp. 331-345.
Montgomery, D.C. 2005. Design and Analysis of Experiments. Sixth edition, Wiley, New
York, NY.
Older, G.J. and Suri, R., 1980. Time Optimal of Part-Routing in a Manufacturing System
with Failure Prone Machines. In: Proceedings of the 19th IEEE Conference on
Decision Control, Vol. 5, pp. 722-727.
98

Presman, E., Sethi, S., and Zhang, Q., 1995. Optimal feedback production planning in a
stochastic N-machine flowshop. Automatica, Vol. 31, No. 9, pp. 1325-1332.
Pritsker, A.A.B. and O’Reilly, J.J. 1999. Simulation with Visual SLAM and Awesim, second
edition. Wiley, West Lafayette, IN.
Rezg, N., Xie, X., Mati, Y., 2004. Joint optimisation of preventive maintenance and
inventory control in a production line using simulation. International Journal of
production Research, Vol. 42, No. 10, pp. 2029-2046.
Rishel, R., 1975. Control of Systems With Jumps Markov Disturbances. IEEE Transactions
on Automatic Control, pp. 241-244
Rossetti, M.D. 2010. Simulation Modeling and Arena. John Wiley and Sons, pp. 573.
Sethi S.P. and Zhang Q., 1994. Hierarchical Control Decision Making in Stochastic
Manufacturing Systems. Birkhäuser Boston, p. 418.
CHAPITRE 5

ARTICLE 3: OPTIMAL LOCKOUT/TAGOUT, PREVENTIVE


MAINTENANCE, HUMAN ERROR AND PRODUCTION POLICIES OF
MANUFACTURING SYSTEMS WITH PASSIVE REDUNDANCY

Behnam Emami-Mehrgani1, Sylvie Nadeau1, Jean-Pierre Kenné1


1
Départment of Mechanical Engineering, École de technologie supérieure, Montreal
(Quebec) H3C 1K3, Canada
This chapter has been submitted as an article for publication in Reliability Engineering &
System Safety. Submitted on March 7, 2012.
Submission Confirmation Ms. Ref. No. : RESS-D-12-00096

Résumé

Cet article est différent par rapport aux autres articles sur l’intégration de C/D. Ce travail
prend en considération le C/D et la maintenance préventive de façon conjointe pour un
système manufacturier constitué de deux machines non-identiques sous forme redondance
passive. Les machines sont sujettes à des pannes, à des réparations aléatoires et à des
activités de maintenance préventive avec ou sans erreur humaine. Les variables de décision
sont les taux de production des machines et le taux de maintenance préventive avec ou sans
erreurs. Les variables de décision influencent le niveau des inventaires et la capacité du
système manufacturier. Le mode de la machine peut être classifié comme étant en opération,
en réparation ou en maintenance préventive avec ou sans erreur. La capacité du système est
décrite par une chaîne de Markov non-homogène. Dans le modèle proposé, le taux de
défaillance du système manufacturier dépend de son âge, ce qui signifie que la politique de
maintenance préventive dépend de l'âge des machines. L’objectif de cette contribution est
divisé en deux (2) volets : 1) trouver les variables de décision permettant de réduire les coûts
totaux de production, comprenant les coûts d’inventaire, de pénurie et de maintenance sur un
horizon de planification infini. 2) vérifier l'influence de l'erreur humaine sur le C/D ainsi que
100

l'activité de maintenance préventive. Des méthodes numériques sont utilisées pour résoudre
les conditions d’optimum et obtenir les politiques optimales qui minimisent le coût sur un
horizon infini. Des exemples numériques et analyse de sensibilité sont présentés pour
illustrer l'utilité de l'approche proposée.

Abstract

The analysis of the optimal production and preventive maintenance with lockout/tagout
planning problem for a manufacturing system is presented in this paper. The considered
manufacturing system consists of two non-identical machines in passive redundancy
producing one type of part. These machines are subject to random breakdowns and repairs.
This paper is different compared to other research projects on preventive maintenance and
lockout/tagout. The influence of human error on lockout/tagout as well as on preventive
maintenance activities are presented in this paper. The decision variables are the production
rate and preventive maintenance rate with lockout/tagout which can influence the inventory
levels and the system's capacity. The system capacity is described by a finite-state Markov
chain. The aim of the study is to minimize production, inventory, backlog and maintenance
costs over an infinite planning horizon; in addition, it aims to verify the influence of human
reliability on the inventory levels for illustrating the importance of human error during the
maintenance and lockout/tagout activities. In this paper, the preventive maintenance policy
depends on the machine age. For the considered manufacturing system the optimality
conditions are provided, and numerical methods are used to obtain machine age-dependent
optimal control policies (production and preventive maintenance rates with lockout/tagout).
Numerical examples and sensitivity analysis are presented to illustrate the usefulness of the
proposed approach.

Keywords: Lockout/tagout, Preventive maintenance, Human error, Production control,


Passive redundancy
101

5.1 Introduction

Production systems optimization is an interesting topic for researchers as well as industry. To


increase the availability of machines at the functional level, several maintenance strategies
have been developed (Lugtigheid et al., 2008; Nahas et al., 2008). The problem of optimally
controlling production rates in a manufacturing system has been widely discussed in the
scientific literature (Hajji et al., 2009; Kenné and Gharbi, 2008; Kenné et al., 2007). In this
view, Rishel (1975) developed the optimal conditions to obtain the optimal solution using
dynamic programming. Based on the formalism of Rishel (1975), Sethi et al. (2002) modeled
the stochastic control production planning for a flexible manufacturing system (FMS) subject
to random failures, which allowed them to obtain the dynamic programming equation of
optimal control. In this way, Sethi et al. (2002) modeled the uncertainties of manufacturing
system by homogeneous Markov processes and determined the production policy in which
the inventory and backlog cost are minimized according to the production rate. The
homogeneous process assumes that state transition rates are constant. This is not actually
applicable in the manufacturing system in which the failure rate depends on the machine age
(Boukas and Haurie (1990)).

Akella and Kumar (1986) modeled a manufacturing system which was consisted of a
machine producing one part type with homogeneous Markov chain. The authors found an
optimal control policy by using Hamilton-Jacobi-Bellman (HJB) equations which is called
Hedging point policy (HPP), (see Dehayem Nodem et al. (2008) for the details). For a
manufacturing system composed of two machines producing one part type, Boukas and
Haurie (1990) showed that the value function that performs the optimal cost must satisfy a
set of differential equations called Hamilton-Jacobi-Bellman (HJB). In the complex
manufacturing systems which optimal conditions are described by the HJB equations,
obtaining an analytical solution of HJB equations remains impossible. The numerical
solution of the HJB equation is a challenge which was considered insurmountable in the
scientific literature. Although there is no analytical solution to HJB equations, Yan and
Zhang (1997) found a solution for the stochastic optimal control problem. Yan and Zhang
used a numerical method based on the Kushner approach (Kushner and Dupuis,1992) for a
102

manufacturing system producing several types of parts. These authors showed that the
obtained control policy is optimal.

Based on the resolution method and the results of Yan and Zhang (1997) several authors
have combined numerical approach with experimental approach based on simulation in order
to extend the concept of non-Markov processes. Furthermore, sometimes it is not possible to
represent failures and repairs of machines by homogenous Markov processes. This case has
been studied by Kenné and Nkeungoue (2008). In this work, authors took into account that
the probability of equipment failures increases according to the machine age. Their work has
been shown that the hedging point policy was optimal for a manufacturing system consisting
of one machine producing one part type. As the failure rate depends on the machine age, the
manufacturing system was modeled by non-homogeneous Markov chain. Since the machines
are subject to random breakdowns and repairs, manufacturing systems operate in a stochastic
environment. Predicting and controlling certain events are possible while others occur
randomly, and are consequently beyond the control of manufacturing systems (Gershwin,
2002). An occupational accident can be observed as a disturbance, often with much more
severe effects than the breakdown itself. These occupational hazards will cause a lot of
damage and prevent the company to perform its tasks.

Occupational safety researchers worldwide have carried out studies confirming the
importance of monitoring and controlling undesirable incidents or accidents during
maintenance procedures (Charlot et al., 2006). An important question that arises is to know:
How can one find an appropriate solution to optimize production while guaranteeing the
safety of workers? One possible answer is lockout/tagout. It consists of locking a machine
with a padlock to discharge all sources of residual energy (hydraulic, electrical, etc.) in order
to avoid the premature starting of equipment throughout an intervention. However, the
current use of lockout /tagout has several shortcomings such as lack of validation of its
feasibility and profitability. There is always a risk during intervention on machines that are
down. The risk of human error during these interventions has a direct impact on the
availability of manufacturing systems and can lead to occupational incidents or accidents.
103

Protection devices are often absent or bypassed during these interventions, which explains
the statistics of accidents remain high in this sector (Chinniah and Champoux, 2008).

In order to address these problems, Emami-Mehrgani et al. (2011) considered an analytical


model combining lockout/tagout, production and corrective maintenance policies for a
passive redundancy system, consisting of two non-identical machines. In this work, the
authors demonstrated clearly that passive redundancy optimized production and maintenance
costs while increasing the security level. Another work integrating lockout/tagout into
operational risk in production control is proposed by Emami-Mehrgani et al. (2012). The
authors considered in Emami-Mehrgani et al. (2012) a manufacturing system consisting of
three machines (two machines with passive redundancy, and one in series with the previous
ones) producing one part type. Their work has confirmed that it is possible to integrate a
passive redundancy system in a production line in order to reduce the production cost. This
also leads to into higher free space-time and minimizes the possibility of circumvention of
protection devices or retraction of lockout/tagout procedures for the machines that are under
repair.

The original nature of works presented in (Emami-Mehrgani et al. (2011) and Emami-
Mehrgani et al. (2012) has allowed them to perceive the importance of human error during
maintenance and lockout/tagout activities. Therefore, this research project takes its
originality from human error modeling in dynamics of manufacturing system and also verify
the influence of human error during maintenance and lockout/tagout activities on the optimal
safety stock levels. In this paper, human error is considered as improper operation which
increases mean preventive maintenance and lockout/tagout time. Hence, control of the
lockout/tagout procedure is integrated into production and maintenance planning. The
objectives of this research project are reached by controlling the production rate, preventive
maintenance rate with lockout/tagout for a manufacturing system consisting of two non-
identical machines in passive redundancy producing one type part. A stochastic dynamic
programming problem is formulated for the considered manufacturing system. Generally
obtaining the optimal solution analytically is difficult, since the corresponding optimality
104

conditions are represented by a set of not easy to solve coupled non-linear partial differential
equations called HJB. In this paper, the numerical methods are used in order to find a
solution of the problem.

This paper is organized as follows: Section 5.2 introduces the notation and assumptions. The
model of the problem under consideration states in Section 5.3. A numerical example and a
sensitivity analysis are provided in Section 5.4. The paper is concluded in Section 5.5.

5.2 Assumptions and notations

This paper incorporates the following assumptions.

5.2.1 Assumptions

1. The corrective and preventive maintenance are carried out with lockout/tagout.

2. The machine becomes new after each corrective maintenance operation.

3. The machine failure rate is a continuous function of its age. It remains undisturbed by
preventive maintenance.

4. The age of machine is rest to zero after each preventive maintenance operation.

5. The preventive maintenance can be achieved with human error or without human error.

6. The mean preventive maintenance time without human error is shorter than mean
preventive maintenance time with human error.

7. The production rate of main machine is upper than the standby machine.

8. The main machine and the standby machine produce the same type of parts

9. The main machine returns to production immediately after each repair and the standby
machine stands idle.

Assumption 5 takes into account the influence of human reliability on the maintenance
activity.
105

Assumption 9 is a classical assumption in passive redundancy system. It is due to the nature


of a passive redundancy system.

5.2.2 Notations

The following notations are used in the rest of this paper:

x(⋅) : inventory/backlog level of finished product

c+ : holding cost incurred on finished product

c− : backlog cost incurred on finished product

cα : cost incurred for the operation of the machine under repair at mode α

cr1 : corrective maintenance cost of main machine M

crs : corrective maintenance cost of standby machine S

ctagout : lockout/tagout cost

u1 : production rate of main machine M

u1max : maximal production rate of main machine M

us : production rate of standby machine S

usmax : maximal production rate of standby machine S

24
w42 : preventive maintenance rate without human error

24 max
w42 : maximal preventive maintenance rate without human error

34
w42 : preventive maintenance rate with human error

34max
w42 : maximal preventive maintenance rate with human error

g (⋅) : instantaneous cost

J (⋅) : total cost


106

v (⋅) : value function

ρ: discount rate

d: demand rate

q14 : repair rate of standby machine S

q23 : preventive maintenance rate without human error to preventive maintenance rate

with human error of standby machine S

q24 : preventive maintenance rate without human error of standby machine S

q34 : preventive maintenance rate with human error of standby machine S

q54 : repair rate of main machine M

5.3 Problem statement

This section develops a manufacturing system for two machines that are not identical in
passive redundancy and producing one type of part. These machines are prone to random
breakdowns and repairs. The machine could be shut down and locked for preventive
maintenance. Repairs will be carried out for machines breakdowns in the context of a year-
to-year operation of the system. The system is shown in Figure 5.1. In the system under
consideration, the main machine operates any time. The other machine called the cold-
standby redundancy goes online immediately when the main machine fails. Then, the
maintenance service is triggered to restore the failed machine. After the failed machine is
restored, it will be returned into the standby position. The main machine has two states:
ξ1 (t ) = 1 if the main machine M is operational and ξ1 (t ) = 2 if the main machine M is
under repair. The standby machine has five states: ξ s (t ) = 1 if the standby machine S is

operational, ξ s (t ) = 2 if the standby machine S is under repair, ξs (t ) = 3 if the standby

machine S is under preventive maintenance without human error, ξ s (t ) = 4 if the standby


107

machine S is under preventive maintenance with human error and ξs (t ) = 5 if the standby

machine S is at time-off. Hence, the dynamics for a manufacturing system with two
machines in passive redundancy is in a hybrid state which consists of a continuous state x (t )

and a discreet state ξ (t ) = (ξ (t )1 , ξ (t ) s ) as follows:

a) Continuous state: The difference between cumulative production and demand denotes as
x (⋅) .The surplus x (⋅) can be positive (i.e., inventory costs c + are thus charged) or negative

(i.e. backlog costs c − are thus charged). The input rate to M (main machine) and S (standby

machine) denotes ui (t ) with (i = 1, S) .

Figure 5.1 Passive redundancy system producing one part type

The dynamics of the system can be written as follows:

dx
= ui (⋅) − d , x (0) = x,
dt
with i = 1, S
da
= f (u (⋅)), a (0) = a,
dt
a (T ) = 0. (5.1)

Where f (ui (⋅)) is an increasing function of the machines production rate which represents

the machine aging.


108

b) Discreet state: Consistent with assumptions set out in previous section, the operational
mode of the whole system can be described by a random vector ξ (t ) = (ξ1 (t ), ξ s (t )) taking

values in Β = {1, 2, 3, 4, 5} . Assuming that in practical terms:

- standby machine cannot be at time-off when the main machine is down;


- standby machine cannot be sent to preventive maintenance, knowing that the main machine
is down and vice versa;
- both machine (main and standby) cannot be sent to preventive maintenance at the same
time;
- both machines (main and standby) cannot be broken down simultaneously (with a good
maintenance plan).

These practical considerations allowed us to describe the dynamics of the system in 5 modes
without loss of generality. Hence, for the passive redundancy ξ (t ) can be expressed as
follows:
1 M is operational and S is under repair;
2 M is operational and S is under preventive maintenance without human error;

ξ (t ) =  3 M is operational and S is under preventive maintenance with human error;
4 M is operational and S is at time-off;

5 M is under rapair and S is operational.
109

Figure 5.2 State transition diagram

The transition rate matrix of the stochastic processes ξ (t ) is denoted by Q such that
Q = {qαβ } , with qαβ > 0 if α ≠ β and qαα = − β ≠α qαβ , where α , β ∈ Β.

The transition probabilities associated to the transition rate qαβ are expressed as:

qαβ (⋅)δt+0(δt) if α≠β,



p ξ(t +δt) = β ξ(t) =α =  (5.2)
1+q (⋅)δt+0(δt) if α=β.
 αβ

The transition rate matrix Q is expressed as follows:

 q11 0 0 q14 0
0 q22 q23 q24 0 

αβ
Q ( w42 )= 0 0 q33 q34 0 , (5.3)
 αβ 
 q41 w42 0 q44 q45 
 0 0 0 q54 q55 
110

Where:
αβ 24
w42 = w42 preventive maintenance rate without human error,
αβ 34
w42 = w42 preventive maintenance rate with human error.

The set of admissible decisions at mode α (t ) and control policies (control variables) at mode
α (t ) :
((u1(⋅), us (⋅), w4224 (⋅), w4234 (⋅)) ∈R4 ,
 max max

0 ≤ u1(⋅) ≤ u1 ,0 ≤ us (⋅) ≤ us ,
Γ(α)  24min 24 24max , (5.4)
w
 42 (⋅) ≤ w 42 (⋅) ≤ w42 (⋅), 
w4234min (⋅) ≤ w4234 (⋅) ≤ w4234max (⋅). 

max
In equation (4), u
1
is the maximal production rate of the main machine, u smax is the
24min
maximal production rate of the standby machine, w42 is the minimal preventive
24 max
maintenance rate without human error , w42 is the maximal preventive maintenance rate
34min
without human error, w42 is the minimal preventive maintenance rate with human error and
34 max
w42 is the maximal preventive maintenance rate with human error. Γ (α ) denotes the set of
all admissible controls.

24 34
The control problem consists in finding an admissible control law u (⋅) = (u1 , us , w42 , w42 ) that

minimizes the cost function J (⋅) given by:

∞ −ρ t 
J ( a, x,α,u1,us , w4224, w4234 ) = E  e g( a, x,α,u1,us , w4224, w4234 ) dt x( 0) = x, ξ ( 0) = a, a(0) = 0, (5.5)
0 

Where ρ is the discount rate and g(x, α , ⋅) = c + x + + c − x − + cα is the instantaneous cost, c + ,

c − and cα , being the cost per unit to produce parts for inventory, backlog as well as
intervention cost on the machine, respectively.

x + = max(0, x) , x − = max(− x, 0)
and
111

c α = ( c r1 + ctagout )Ind {α = 1} + ( c pm + ctagout )Ind {α = 2}


+ ( c pm + ctagout )Ind {α = 3} + ( c rs + ctagout ) Ind {α = 5}

With:
1 if Θ(⋅) is true 
Ind {Θ(⋅)} =  
0 otherwise 
For a given proposition Θ(⋅) .

As in Akella and Kumar (1986), the manufacturing system considered has a sufficient
capacity to ensure its feasibility, which means the average capacity is not less than the
demand rate.

Letν ( a, x,α ) denote the value function or minimum discounted cost for equations (5.5) as
expressed in the following equation:

ν ( a, x,α) = min
,w42 )∈Γ( x,α)
24 34
(u1,us ,w42
( 24
42
34
)
J a, x,α,u1, us, w , w42 , ∀α ∈Β (5.6)

In Appendix 5.A, the properties of the value function v(⋅) given by equation (5.6) are
presented . It is shown that the value function v(⋅) given by (5.6) should satisfy a set of
partial differential equations known as the Hamilton-Jacobi-Bellman (HJB) equations.

5.4 Numerical example and results analysis

In this section, we present a numerical example for a manufacturing system consisting of two
non-identical machines in passive redundancy. The system capacity is described by a five
Markov process with states ξ (t ) ∈ Β = [1,2,3,4,5] . The generator matrix Q(⋅) described by
equation (5.3) is explicitly defined as follows:

 −q14 0 0 q
14  0
 
 0 −( q + q ) q q 0 
23 24 23 24
αβ 
Q( w ) =  0 0 −q q 0 ,

42 34 34
 αβ αβ 
 q41 w
42
0 −(q + w + q (a)) q (a) 
41 42 45 45
 
 0 0 0 q
54
q55 

Where:
112

αβ 24
w42 = w42 preventive maintenance rate without human error,
αβ 34
w42 = w42 preventive maintenance rate with human error.
and
−( At ×a (t )2 ) 
q45 ( a (t )) = A45∞ 1 − e , (5.7)
 

or given constants A45∞ and At . The transition rate q45 (⋅) given by equation (5.7), illustrates

the impact of a machine age on its dynamics as in Kenné and Nkeungoue (2008). The mean
time between failures (MTBF) is machine-age dependent and is given by:

1 1
MTBF(a ) = with MTBF(∞) = ∞ .
q45 (a ) A45
The discrete dynamic programming equation (5.A.4) gives the following five equations:

 u1 − d h 
 (ν ( x+hx ,α,1) k + +ν h ( x-hx ,α,1) k − )
−1
 hx 
 u1 − d Ku   Ku h 
ν h (α,x,1) = min ρ + + + q14  + (ν ( x+hα ,x,1) + g ( x,α,1) , (5.8)
,w42 )∈Γ(1)
24 34
(u1 ,us ,w42
 hx ha   ha 
+q ν h ( x,α,4) 
 14 
 

 u1 − d h 
 (ν ( x+hx ,α,2) k+ +ν h ( x-hx ,α,2) k− )
−1
 hx 
 u1 − d Ku   Ku h 
ν h (α,x,2) = min ρ + + + q23 + q24  + (ν ( x+hα ,x,2) + g ( x,α,2) , (5.9)
(u1,us ,w42 ,w42 )∈Γ( 2)
24 34
 hx ha   ha 
+q ν h ( x,α,3) + q ν h ( x,α,4) 
 23 24

 

 u1 − d h 
 (ν ( x+hx ,α,3) k+ +ν h ( x-hx ,α,3) k− )
−1
 hx 
 u1 − d Ku   Ku h 
ν h (α,x,3) = min ρ + + + q34  + (ν ( x+hα ,x,3) + g ( x,α,3) , (5.10)
,w42 )∈Γ( 3)
24 34
(u1 ,us ,w42
 hx ha   ha 
+q ν h ( x,α,4) 
 34 
 
113

 u1 −d h 
 (ν ( x+hx ,α,4) k+ +ν h ( x-hx,α,4) k−)
−1
 hx 
 u1 −d Ku   Ku h 
ν h (α,x,4) = min ρ + + + q41 + wαβ  + ν
( ( x+hα ,x,4 ) + g ( x,α,4) , (5.11)
,w42 )∈Γ( 4)
24 34
(u1,us ,w42
 hx ha 42
  ha 
+q ν h ( x,α,1) + wαβν h ( x,α,2) 
 41 42

 

 us − d h 
 (ν ( x+hx ,α,5) k+ +ν h ( x-hx ,α,5) k− )
−1
 hx 
 us − d Ku   Ku h 
ν h (α,x,5) = min ρ + + + q54  + (ν ( x+hα ,x,5) + g ( x,α,5) + , (5.12)
,w42 )∈Γ( 5)
24 34
(u1,us ,w42
 hx ha   ha 
q ν h ( x,α,4) 
 54 
 

The following computational domain is used:

Gaxh = {( x, a ) : −5 ≤ x ≤ 25; 0 ≤ a ≤ La } , in which hx = 2, ha = 3and La = 150.

La should be smaller than the age limit of the machine ( Lmax ) with respect to the feasibility

conditions.

The characteristics of main and standby machines must be in such a way that respects the
feasibility of system.

The system is considered feasible if:

π (⋅)uimax (⋅) ≥ d (5.13)

Where the limitation probabilities can be ascertained from the following equation for a
system conforming to a Markov process:

π (⋅)Q(⋅) = 0,

π = 1
with:
114

π (⋅) : Limiting probabilities

Q (⋅) : Transition matrix rates

π (⋅) = (π1 , π 2 , π 3 , π 4 , π 5 ) representing the vector of limiting probabilities from modes 1 to 5.


Using the formula (5.13), we ascertained that the system is not feasible for an age of 287, that
means Lmax = 286.

Figure 5.3, illustrates the failure rate of the main machine for each value of its age with
values of A45∞ = 0.1and At = 1× 10−5 chosen to obtain a failure probabilities trajectory

according to the machine age as in Kenné and Nkeungoue (2008).

Figure 5.3 Age-dependent failure rate of the main machine


The parameters for our case study appear in Table 5.1.
115

Table 5.1 Parameters of the numerical example

u1max u smax 24min


w42 24 max
w42 34min
w42 34 max
w42 q14 q23 q24 q34 q54
-6 -6 -5
0.32 0.28 10 0.06 10 0.06 10 0.02 0.083 0.125 0.053
d ρ c +
c −
cr cmp ctagout
0.25 0.01 1 30 7900 400 100

Please note that these parameters have been taken so that the system will be feasible. Further,
we will check the influence of these parameters by sensitivity analysis. The policy iteration
technique solves equations (5.8)–(5.12) for optimal conditions. Recall that the control
policies are obtained from the numerical resolution of the optimality conditions given by
equations (5.8)–(5.12). The structure of production policies (u1 and us ) states that: if the stock
level is lower than a threshold level, then produce at maximum rates; else if the stock level is
upper than a threshold level produce nothing; otherwise produce at the demand rate. Such
structure in the control literature is called hedging point policy (HPP).

24 34
The results obtained for the control variables u1 , us , w42 and w42 of a passive redundancy
system are given in Figures (5.4)-(5.8) for illustration purposes.

The machine-age-dependent threshold value, for data presented in Table 1, is defined using
the switching trend illustrated in Figures (5.4)-(5.8). Hence, the production policy is given by
as follows:
u max (α ) if x(⋅) < ψ (a ),
 i
ui (a, x, α ) =  d if x(⋅) = ψ (a ), (5.14)
 0 otherwise,

Where:
uimax (α ) = u1max if α = 1, 2,3, 4 , uimax (α ) = usmax if α = 5.
and ψ ( a ) is machine-age dependent function that gives the optimal threshold value for each
value of the machine age. Hence, we obtain from numerical result as follows:
116

 X * (a ) if a (t ) ≥ A* (a ),
ψ (a (t )) =  (5.15)
 0 otherwise.

Figure 5.4 shows the existence of different zones for the production rate of the main
machine. Each of these areas follows a production optimal policy. These different production
policies reflect the machine age and the inventory level. According to this figure, beyond a
certain age the threshold becomes more important. The policy suggests setting the production
rate of the machine to its maximal value when the current stock level is under an age-
dependent threshold value. In this area the age leads to frequent failure. On the other hand,
the production rate of the machine sets to the demand rate when the current stock level is
equal to an age-dependent threshold value. Finally the production rate of the machine sets to
zero when the current stock level is larger than an age-dependent threshold value.

Figure 5.5 illustrates the same structure as Figure 5.4, but the policy suggests keeping more
inventories by increasing the machine age, because the preventive maintenance has been
done with the human error. Figure 5.6 plots the same policy for standby machine production
rate as main machine production rate, but this figure is illustrated to have more inventories
according to machine age. Recall that the standby machine production rate is lower than the
main machine production rate.
117

Figure 5.4 Main machine production rate at mode 2

Figure 5.5 Main machine production rate at mode 3


118

Figure 5.6 Standby machine production rate at mode 5

The preventive maintenance policy with human error and without human error is plotted in
Figures 5.7 and 5.8. Figure 5.7 shows the standby machine preventive maintenance rate
without human error, plotted in this figure, divides the computational domain (x,a) into two
regions . Hence, the preventive rate is set to its maximal value for backlog situation and to
zero for large stock levels. For significant stock levels, the zone in the domain (x,a) where
the preventive maintenance is set to its maximal value increases with the machine age. Figure
5.8 illustrates the same structure as Figure 5.7, but the policy suggests keeping more
inventories by increasing the machine age in order not to penalize the demand, for the reason
that the preventive maintenance has been done with the human error. The corresponding
optimal policy, similar to the previous policy, has a bang bang structure and is given by
following equation:
119

αβ
 wmax αβ if x (⋅) < B* ( a ),
w42 =  42 (5.16)
min αβ
 w42 otherwise.
Where:
αβ 24
w42 = w42 preventive maintenance rate without human error,
αβ 34
w42 = w42 preventive maintenance rate with human error.
and B* ( a ) is machine-age-dependent function that gives the optimal stock level at which
it is necessary to switch the preventive maintenance from w αβ αβ
42 to w 42 for given machine age.

Figure 5.7 Standby machine preventive maintenance rate without human error at mode 2
120

Figure 5.8 Standby machine preventive maintenance rate with human error at mode 3

The influence of the backlog cost on the production threshold according to the machine age
as well as preventive maintenance cost according to the machine age for main and standby
machine are presented in Figures 5.9 and 5.10.

Figure 5.9 Trend of threshold value of main machine according to machine age
121

WOHE*: Without human error

WHE*: With human error

Indeed, Figure 5.9 illustrates that if the production unit is still in its youth phase, it is not
essential to store large inventory levels. In this phase, the failure rate is low and the
production unit has a high reliability and it is able to meet the demand.
As one goes along the machine age increases, storage of finished products inventories are
suggested to meet the demand. The quantity of finished products inventories depends on the
machine age. When the machine becomes older, the higher level of inventories is needed.
Hence, the manufacturing system is preparing for a possible failure whose frequency
increases with age. It should be also remarked that if the backlog cost increases, the level of
stock must be increased for the same machine age to avoid a possible shortage. This increase
is more important, if the maintenance technician makes errors during the maintenance
activity.

Figure 5.10 Trend of preventive maintenance rate of standby machine according


to machine age
122

WOHE*: Without human error

WHE*: With human error

The Figure 5.10 shows the preventive maintenance rate that is based on the machine age and
inventory level. This figure can be divided into two main areas I and II. In Zone II, the
production policy suggests not to send the machine in preventive maintenance. Indeed, in this
area, the production unit is still new and the probability of failure is almost zero. As a result,
the manufacturing system is able to meet the demand without fear of failure. Therefore, it is
unnecessary to perform preventive maintenance. This maintenance is appropriate only in
zone I, as shown in Figure 5.10. Indeed, in this area the production unit is aging and the
failure rate increases. Perform preventive maintenance becomes necessary to increase the life
of the production system. Therefore, the manufacturing system must ensure a certain
inventory level to meet the demand. In addition, if the cost of preventive maintenance
increases, the frequency of preventive maintenance must be reduced. It is also observed that
if the maintenance technician makes a mistake, a larger inventory level for the same cost of
preventive maintenance must be stored in order to meet the demand. Moreover, because of
this anomaly the occurrence probability of an incident or accident will be increased.

Optimal costs of production and preventive maintenance including lockout/tagout procedures


for two types of preventive maintenance (without human error and with human error) may be
determined using the analytical model presented in this paper. In this paper the human error
increases repair time which is exponentially distributed. If the repair-time is not
exponentially distributed (non-Markovian case), an extension of human error concept
developed in this paper can be used. The numerical approach demonstrates that the resulting
policy is optimal and enhances machine availability for manufacturing system under
consideration. The control policy for manufacturing system considered is based on an
extension of the hedging point structure. Without limiting in any way the generality of this
proposal, the extension of this model might be adapted to different industrial sectors under
certain conditions.
123

5.5 Conclusion

This paper sought to verify the influence of human error on the preventive maintenance
activity with lockout/tagout. In this paper, the control policy is based on an extension of the
hedging point structure. Based on the numerical solution obtained, a parameterized near-
optimal control policy was derived. The proposed approach using a numerical example and a
sensitivity analysis are illustrated and validated. This work demonstrates clearly that human
error during maintenance activities can increase the production cost while reducing the safety
of workers. In order to avoid this problem, the maintenance technicians must be better trained
and also the production cadence must not be increased brutally. Because the maintenance
technician can make a mistake and the occurrence probability of an incident or accident will
be increased. Furthermore, a larger inventory level must be stored in order to meet the
demand during the maintenance activity. Lastly, a number of conditions must be met to make
effective use of the model presented in this paper.

Acknowledgments

We gratefully acknowledge for financial support of this work from the following agencies:
Natural Sciences and Engineering Research Council of Canada (NSERC); Réseau de
Recherche en Santé et en Sécurité du Travail du Québec (RRSSTQ); Fonds de la Recherche
en Santé – Québec (FRSQ); Fonds Québécois de la Recherche sur la Société et la Culture
(FQRSC); Fonds Québécois de la Recherche sur la Nature et les Technologies (FQRNT) ,
Institut de Recherche Robert-Sauvé en Santé et en Sécurité du Travail (IRSST) and Équipe
de Recherche en Sécurité du Travail (EREST).
APPENDIX 5.A OPTIMAL CONDITIONS AND NUMERICAL APPROACH

The properties of the value functions and the Hamilton-Jacobi-Bellman equation (HJB) are
presented in this section. These equations describe the optimality conditions for a
manufacturing system consisting of two non-identical machines in passive redundancy
producing one type of part. The HJB equation is adapted to the optimal control problem
considered as follows:

 u(α) −d ∂ ν a, x,α + Ku ∂ ν a, x,α g α, x, ⋅ 


( ) ( )
∂x ∂a
( ) ( )
ρν ( a, x,α) = min  , ∀α,β ∈Β (5.A.1)
,w42 )∈Γ( x,α)
24 34
(u1,us ,w42
+  qαβν ( x,β) 
 α≠β 

With:
u(α ) = u1 if α = 1, 2,3, 4 and u (α ) = us if α = 5 .

The optimal control policy (u1* , us* , w42


24* 34*
, w42 ) denotes a minimizer over Γ(α ) of the right-

hand side of equation (5.A.1). This policy is consistent with the value function obtained in
Equation (5.6). The optimal control policy therefore rests in solving Equation (5.A.1).
Obtaining an analytical solution of equation (5.A.1) is roughly impossible. The numerical
solution of the HJB equation (5.A.1) is a challenge considered insurmountable in the
scientific literature.

The numerical method for solving the optimality conditions is presented in this section. This
method is based on the Kushner approach (see Kushner and Dupuis (1992) for details). The
main idea behind this approach consists of using an approximation scheme for the gradient of
the value function ν ( a, x,α ) . Let hx and ha denote the length of the finite difference interval

of the variables x and a respectively. Using hx ,ν ( x,α ) is approximated by ν h ( a, x,α ) and

ν x ( a, x,α ) is approximated as follows:


125

1

ν x ( a, x,α ) × (u(α) − d ) =  h
(ν h ( a, x + hx,α ) −ν h ( a, x,α )) ×(u(α) − d) 
if ( u(α ) − d ) ≥ 0 

, (5.A.2)
1
 h (ν h ( a, x,α ) −ν h ( a, x-hx,α )) ×(u(α) − d) otherwise 

With:

u(α ) = u1 if α = 1, 2,3, 4 and u (α ) = us if α = 5 .

Using ha , ν a ( a, x,α ) is approximated as follows:

ν a ( a, x,α ) × f (u) =
1 h
h
( )
ν (α , x + ha , x ) −ν h ( a, x,α ) × f (u), (5.A.3)

With approximations given by equations (5.A.2) and (5.A.3) and after some manipulations,

the HJB equations (5.A.1) can be rewritten as follows:


 −1
u(α ) − d Ku   u(α ) − d h
ν ( a, x,α ) =
h
min
, w42 )∈Γh (α )
24 34  ρ + qαα +
hx
+  ⋅
ha   hx
(ν ( a, x + hx ,α ) K +
(u1 ,us , w42

Ku  
+ν h ( a, x − hx ,α ) K − )+ vh (a, x + ha ,α ) + g(a, x,α ) + Q.ν h (a, x,α ) , (5.A.4)
ha  
With:

u(α ) = u1 if α = 1, 2,3, 4 and u (α ) = us if α = 5 .

Where Γ (α ) is the discreet feasible control space and the other term used in equation

(5.A.4) is defined as:

1 if (u (α ) − d ) ≥ 0,
K+ = 
0 otherwise,

1 if (u (α ) − d ) < 0,
K− = 
0 otherwise,
With:

u(α ) = u1 if α = 1, 2,3, 4 and u (α ) = us if α = 5 .


126

The equation presented in (5.A.4) can be interpreted as the infinite horizon dynamic
programming equation for a discreet-time, discreet-state decision process that addresses
problems confronted in optimizing output and controlling maintenance. (see Boukas and
Haurie (1990) for details).
The next theorem shows that value function ν h ( a, x,α ) is an approximation to ν x ( a, x,α ) for

a small step size hx and ha .

Theorem

Let v h (a, x, α ) denote a solution to HJB equation (5.A.4). Assume they are constants C g and

K g as follows:
k
0 ≤ v h ( x, α ) ≤ Cg (1 + x g
)

Then,
lim vh (a, x, α ) = v(a, x, α )
h→0

Proof

The proof of this theorem is given in Emami-Mehrgani et al. (2011).


127

REFERENCES

Akella, R. and Kumar, P.R., 1986. Optimal Control of Production Rate in a Failure-Prone
Manufacturing System. IEEE Transactions on Automatic Control AC 31, pp. 116-
12.
Boukas, E.K. and Haurie, A., 1990. Manufacturing flow control and preventive maintenance:
A stochastic control approach. IEEE Transactions on Automatic Control AC 33, pp.
1024–1031.
Charlot, E., Kenné, J.P. and Nadeau, S., 2006. Optimal Production, Maintenance and
Lockout/tagout Control Policies in Manufacturing Systems. International Journal of
Production Economics 107, pp. 435-450.
Chinniah, Y. and Champoux, M., 2008. La sécurité des machines automatisées : Analyse des
risques et des moyens de protection sur une presse à injection de plastique. R-557.
Montréal: IRSST, pp. 93.
Dehayem Nodem, F.I, Kenné J.P., and Gharbi, A., 2008. Joint Optimization of Capacity
Expansion, Production and Maintenance in Stochastic Manufacturing Systems.
Applied Mathematical Science, Vol. 2, No. 5-8, pp.195–212.
Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011. Lockout/tagout and operational risks
in the production control of manufacturing systems with passive redundancy.
International Journal of Production Economics 132, pp. 165-173.
Emami-Mehrgani, B., Kenné J.P., and Nadeau S., 2012. Lockout/tagout and Optimal
Production Control Policies in Failure-prone Non-homogeneous Transfer Lines
With Passive Redundancy. Accepted (January), International Journal of Production
Research
Gershwin, S.B., 2002. Manufacturing Systems Engineering. Prentice-Hall, Englewood Cliffs,
NJ, pp. 267.
Hajji, A., Gharbi, A., Kenné, J.P., and Hidehiko, Y. 2009. Production and changeover control
policies of a class of failure prone buffered flow-shops. Production Planning &
Control, Vol. 20, No. 8, pp. 785-800.
Kenné, J.P., Gharbi, A., and Beit, M., 2007. Age-Dependent Production Planning and
Maintenance Strategies in Unreliable Manufacturing Systems with Lost Sale.
European Journal of Operational Research, Vol.178, pp. 408–420.
Kenné, J.P., Gharbi A, Najid M. 2009. On the Integrated Production, Inventory, and
Preventive Maintenance Problem in Manufacturing Systems with Backorder.
International Journal of Simulation and Process Modelling, Vol.5, N0. 4 pp. 300-
312.
Kenné, J.P., and Nkeungoue, L.J., 2008. Simultaneous Control of production and corrective
maintenance rates of a failure-prone manufacturing system. Applied Numerical
Mathematics; Vol.58, pp.180-194.
128

Kushner, H.J. and Dupuis, P.G., 1992. Numerical Methods for Stochastic Control Problems
in Continuous Time. Springer, New York.
Lugtigheid, D., Banjevic, D., and Jardine A.K.S., 2009. System repairs: when to perform and
what to do?. Reliability Engineering and System Safety; Vol.93, No.4; pp.604–15.
Nahas, N., Khatab, A., Aït-Kadi D., and Nourelfath M., 2008. The extended great deluge
algorithm for the imperfect zaintenance optimization problem of multi-state
systems. Reliability Engineering and System Safety, Vol. 93, No. 11, pp.1658-1672.
Rishel, R., 1975. Control of Systems With Jumps Markov Disturbances. IEEE Transactions
on Automatic Control, pp. 241-244
Sethi, S.P, Yan, H., Zhang, H., and Zhang, Q., 2002. Optimal and hierarchical controls in
dynamic stochastic manufacturing systems: a survey. Manufacturing & Service
Operations Management, Vol.4, No. 2, pp. 133-170.
Yan, H., and Zhang, Q., 1997. A numerical method in optimal production and setup
scheduling to stochastic manufacturing systems. IEEE Transactions on Automatic
Control, Vol.42, pp.1452–1455.
129

CONCLUSION GÉNÉRALE

L’utilisation actuelle du C/D présente certaines lacunes. Un niveau de risque élevé existe
pendant l’intervention sur des machines en pannes. Le risque d’erreur humaine au cours des
interventions présente une probabilité importante d’impact sous forme d’incidents ou
d'accidents du travail chez les maintenanciers et sur la disponibilité du système
manufacturier. Afin de remédier à ces lacunes cette thèse a été développée en trois (3)
parties.

Dans la première partie, l'intégration du MTTLT et de la maintenance corrective pour un


système manufacturier est combinée à la gestion de la capacité de production. Le système
manufacturier est constitué de deux machines non-identiques sous forme redondance passive
produisant un seul type de produit. L'objectif de cette intégration est d'augmenter le niveau de
sécurité des travailleurs ainsi que la disponibilité des machines. Une modélisation a été faite
par chaîne de Markov homogène et une résolution numérique à travers des équations
différentielles d'HJB a conduit à la solution du système étudié. Les diverses analyses de
sensibilité ont été effectuées afin de confirmer la structure des politiques obtenues.

La deuxième partie aborde également le problème d'intégration du MTTLT dans la gestion


de la capacité de production. Le système est modélisé par une chaîne de Markov homogène.
L'influence de contrôle du MTTLT a été présenté pour une ligne de production constituée de
trois machines (deux machines sous forme redondance passive et une troisième machine en
série avec les précédentes) produisant un type de pièce. Afin d’avoir un modèle plus réaliste
pour les industries, les conditions d’optimum proposées sont développées en utilisant une
approche combinée, en se basant sur une combinaison de formalisme analytique, la
simulation, le plan d'expérience et la méthodologie de surface de réponse. Une analyse de
sensibilité a illustré l'utilité de nos résultats.

La troisième partie du travail s'est concentré sur la modélisation du MTTLT et l'erreur


humaine pour un FMS. Le FMS est constitué deux machines non-identiques sous forme
redondance passive produisant un seul type de produit. La théorie de commande est basée sur
130

des modèles mathématiques. En effet, dans cette partie nous avons prévu une modélisation
analytique par chaîne de Markov non homogène, dont la résolution est faite d’une part par la
résolution d’équations différentielles d'HJB et par analyse numérique pour illustrer l'utilité de
l'approche proposée. Compte tenu du fait que l'erreur humaine allonge les temps de
réparation, nous avons montré que l'introduction de l'erreur humaine dans les activités de
C/D ainsi que de maintenance peut engendrer un accident grave et une augmentation de coût
de production.

Dans cette thèse, notre travail a apporté une contribution scientifique significative en
reformulant les modèles mathématiques existant pour intégrer le MTTLT en contexte FMS.
Les résultats de nos travaux ont été confirmés à travers des études par modélisation,
résolution numérique, l'approche combinée et analyse de sensibilité sur des cas de FMS. Ce
travail a confirmé qu'en intégrant le MTTLT dans un système en redondance passive, le
système devient moins vulnérable aux variations des coûts de pénurie, d’inventaire et de
maintenance en satisfaisant la demande en permanence. Cette intégration dans un système
manufacturier en redondance passive a permis de libérer un espace-temps essentiel pour
minimiser les possibilités de contournement des dispositifs de protection ou d’escamotage
des procédures de C/D. Nos travaux ont été conclus par l'exploration de l'impact de l'erreur
humaine sur le C/D ainsi que les activités de maintenance.
131

TRAVAUX FUTURS

À l'issue de cette thèse, l'intégration du C/D dans la gestion de la capacité de production a été
abordée, ensuite l'influence de l'erreur humaine sur les activités de C/D et de maintenance a
été vérifiée. Ces contributions constituent une base solide pour des travaux futurs.

Les résolutions proposées dans cette thèse peuvent être étendus à des systèmes plus
complexes de point vue structure et taille afin d'ouvrir une nouvelle piste de recherche. Nos
résultats de recherche peuvent être utilisés directement sur des cas réels. Cependant, nous
pouvons faire appel à d'autres méthodes de modélisation (semi-Markovien, non-Markovien)
et améliorer la résolution numérique des conditions d'optimum. Ceci permettrait de faire une
étude comparative des résultats trouvés. Nous pouvons élargir cette étude à d’autres modèles
comme M2P2 (deux machines produisant deux types de produit). Dans ces modèles, il faut
également définir une politique optimale de maintenance corrective et préventive. Le système
manufacturier sujet à des demandes aléatoires peut être considéré afin de trouver une
politique optimale de planification de C/D. D'autres prochains développements dans ce projet
de recherche peuvent prendre en considération la réduction de capacité et son influence sur
les activités de C/D et de maintenance ainsi que l'intégration de la notion des taux de rejets
pour le système manufacturier. Cette notion de taux de rejet permettra de mesurer les impacts
réelles de tous les facteurs ayant de l’influence sur le MTTLT, la maintenance et la
production tels que l’équipement, l’environnement et l’humain. Nous pouvons également
prendre en considération la notion de contrôle de la qualité dans l'environnement dynamique
stochastique, ce qui est incontournable dans le processus de transformation (matière
première, encours, produits finis). Par ailleurs, n'oublions pas l'importance de recueillir les
données réelles dans un site industriel et alimenter nos modèles proposés par ces données
réelles afin d'évaluer la performance des politiques obtenues. Élaborer ces différents points
permettrons d'améliorer la production et le bien-être des travailleurs, qui sont les principaux
axes de la gestion des opérations, de l’ergonomie et du génie des facteurs humains.
RÉFÉRENCES

Akella, R. and Kumar, P.R., 1986. Optimal Control of Production Rate in a Failure-Prone
Manufacturing System. IEEE Transactions on Automatic Control AC 31, pp.
116-126.
Badiane, A., 2010. "L'intégration des stratégies de maintenance à la gestion des opérations
des systèmes manufacturiers", Mémoire de maîtrise en génie de la production
automisée, École de technologie supérieure (ÉTS), Montréal, pp.107.
Barlow, R., Frank, P. and Larry, H., 1996. Mathematical theory of Reliability.« SIAM ».
New York, Society of applied mathematics, pp. 258.
Basnet, C. and Mize, J., 1986. An Object-Oriented Framework for Operating Flexible
Manufacturing Systems. Proceedings of the International Conference on Object-
Oriented Manufacturing Systems, pp. 346-341.
Ben-Daya, M., Salih, D., and Abdul, R., 2000. Maintenance, Modeling and optimization.
«Norwell». Massachusetts : Kluwer Academic Publisher, pp.474
Bhattacharya, A.K. and Coleman, J.L., 1994. Linking Manufacturing Strategy to the Design
of a Customized Hybrid Production Control System. International Journal of
Computer-Integrated Manufacturing Systems, vol.7, No.2, pp. 134-41.
Boukas, E.K. and Haurie, A., 1990. Manufacturing Flow Control and Preventive
Maintenance: A Stochastic Control Aproach. IEEE Transactions on Automatic
Control 33, pp. 1024-1031.
Boukas, E.K. and Kenné, J.P., 1997. Maintenance and Production Control of Manufacturing
Systems with Setup. Lecture in Applied Mathematics, Vol .33, pp. 55-70.
Charlot, E., Kenné, J.P. and Nadeau, S., 2006. Optimal Production, Maintenance and
Lockout/tagout Control Policies in Manufacturing Systems. Int. J. Production
Economics 107, pp. 435-450.
Gershwin, S.B., 2002. Manufacturing Systems Engineering. Prentice-Hall, Englewood Cliffs,
NJ, pp. 267.
Gharbi, A. and Kenné, J.P., 2003. Optimal Production Control Problem in Stochastic
Multiple-Product Multiple-Machine Manufacturing Systems. IIE Transactions
35, pp. 941-952.
Gwo-Liang, L., Yen Hung, C. and Shey-Huei, S., 2009. Optimal Economic Production
Quantity Policy for Imperfect Process with Imperfect Repair and Maintenance.
European Journal of Operational Research, Vol. 195, pp. 348-357.
Harichaux, P. and Libert, J., 2003. Ergonomie et Prévention des Risques Professionnels,
L'Environnement Physique du Travail et ses Contraintes, vol. 1, Chiron, pp.192.
Kenné, J.P. and Boukas, E.K, 1998. Optimal Control of Failure prone Manufacturing System
with Corrective Maintenance. In: Proceeding of the 18th IFIP TC7 Conference,
pp. 468-476.
134

Kenné, J.P. and Gharbi, A., 1999. Experimental Design in Production and Maintenance
Control Problem of a Single Machine, Single Product Manufacturing System.
International Journal of Production Research, Vol.37, No.3, pp. 621-637.
Kenné, J.P. and Gharbi, A., 2000. A Production Planning Problem in Manufacturing Systems
with General Failure and Repair Times Distributions. Production Planning &
Control Journal. Vol.11, No.6, pp. 581-587.
Kenné, J.P. and Gharbi, A., 2001. A Simulation Optimization Approach in Production
Planning of Failure Prone Manufacturing Systems. Journal of Intelligent
Manufacturing Vol.12, No.5, pp.421-431.
Kenné, J.P. and Gharbi, A., 2003. Stochastic Optimal Production Planning Control Problem
with Corrective Maintenance. Computers & Industrial Engineering, Vol. 46, No.
4, pp. 865-875.
Kenné, J.P., 2004. Commande Optimal Stochastique Appliquée à la Gestion de Production
d’une Unité de Fabrication. Journal Européen des Systèmes Automatisés (JESA):
pp. 44-64.
Kenné, J.P. and Gharbi, A., 2004. A Simulation Optimization Based Control Policy for
Failure Prone One-Machine, Two-Products Manufacturing System. Computer
and Industrial Engineering, Vol. 46, No.2, pp. 285-292.
Kimemia, J. and Gershwin, S.B., 1983. An Algorithm for the Computer Control of
Production in Flexible Manufacturing Systems. IEEE Transactions on Automatic
Control AC 15. pp. 353-362.
Kushner, H.J. and Dupuis, P.G., 1992. Numerical Methods for Stochastic Control Problems
in Continuous Time. Springer, New York, pp. 473.
Mansour, M., Nadeau, S et Kenné, J.P., 2009. "Management and Planning of Lockout
Operations : The Case of a High Volume Packing Line", Gesellschaft für
Arbeitswissenschaft (GfA), Mars, Dortmund, Allemagne, pp.547-551.
Matsuoka, S. and Muraki, M. 2001. "Computer aided planning for lockout/tagout program"
Japan, Process safety progress, Vol. 20, No.2, pp.130-135.
Monchy, F., 1991. La fonction maintenance. « Masson ». Paris, pp. 452.
Nollet, J., Kelada, J., O.Diorio, M., Handfield, H., Deschamps, I., R.Duguay and C., 1994. La
Gestion des Opérations et de la Production, Une Approche Systématique. 2e
Édition, Gaëtan Morin. pp. 163.
Older, G.J and Suri, R., 1980. Time Optimal of Part-Routing in a Manufacturing System with
Failure prone machines. In: Proceedings of the 19th IEEE Conference on
Decision Control, vol. 5, pp. 722-727.
Presman, E., Sethi, S. and Suo, W., 1997. Existence of Optimal Feedback Production Plans
in Stochastic Flowshops with Limited Buffers, Automatica 33, pp.1899-1903.
Rishel, R., 1975. Control of Systems With Jumps Markov Disturbances. IEEE Transactions
on Automatic Control, pp. 241-244
135

Sader, B.H. and Sorensen, C.D., 2003. Deterministic and Stochastic Dynamic Modeling of
Continuous Manufacturing Systems Using Analogies to Electrical systems. In:
Proceedings of the 2003 Winter Simulation Conference. Vol.2, pp. 1134-1142.
Sethi, S.P., Zhang, Q. and Zhou, X.P., 1992. Hierarchial Controls in a stochastic Two-
machines flowshop with Finite Internal Buffer. In: Procedding of the 31 st IEEE
Conference on decision and Control, pp. 2074-2079.
Stevenson, W et. C. Benedetti. 2006. La gestion des opérations : Produits et services, 2e éd
Montréal : Chenelière/McGaw-Hill, pp.249.
Thesen, A., 1999. Some Simple, But Efficient, Push and Pull Heuristics for Production
Sequencing for Certain Flexible Manufacturing Systems. International Journal of
Production Research, Vol. 37, pp.1525-1539.
ANNEXE 1

INTEGRATING LOCKOUT/TAGOUT WITH OPERATIONAL RISKS: THE


PASSIVE REDUNDANCY CASE

Behnam Emami-Mehregani1,Sylvie Nadeau1, Jean-Pierre Kenné1


1
Départment of Mechanical Engineering, École de technologie supérieure, Montreal
(Quebec) H3C 1K3, Canada
Gesellschaft für Arbeitswissenschaft (GFA)
24-26 March 2010, Darmstadt, Germany

Abstract

In this paper, we consider a production control and occupational safety problem for a failure
prone manufacturing system consisting of two machines (non-identical) in passive
redundancy case. Machines are subject to failures and repairs and can produce one type of
part. Therefore we are concerned by the impact of production control including
lockout/tagout on production cost and corrective maintenance cost for two type of failures in
a flexible manufacturing system (FMS) environment. The aim of this study is to demonstrate
the optimal production cost, backlog, inventory and maintenance cost over an infinite
planning horizon and the impact on occupational risk level when one offers an efficient
planning of lockout/tagout in production control.

1. Introduction

How can one find an appropriate solution to optimize production while guaranteeing the
safety of workers? One possible answer is lockout/tagout. It consists in locking a machine
with a padlock, then to discharge all sources of residual energy (potential, hydraulic,
electrical, etc.) to avoid the premature starting of equipment throughout an intervention. In
the current industrial situation, many managers think wrongfully that planning and realizing
lockout/tagout takes a lot of time. This inactive time of production is considered reducing
firms’ performance. This manuscript demonstrates that it is possible to improve the
138

performance of flexible manufacturing system (FMS) as for fabrication costs and


occupational risk level by offering an efficient planning of lockout/tagout.

This paper is organized as follows:

• firstly, we verify the influence of lockout/tagout activities and time to repair the
failure of a machine in a flexible manufacturing system (FMS) environment. In this
research, it is assumed that the machine produces one type of part.
• secondly, we integrate the lockout/tagout in production control. We consider
lockout/tagout for a system in passive redundancy which consists of two (2) machines
(identical and non-identical). Afterword, we propose, by using numerical methods, a
production policy for a system constituted of two (2) machines in passive
redundancy.
• finally, we make a comparison between the results in this paper and the results of
Charlot et al. (2006).

To reach our targets, we follow these steps:

• develop an analytical model of lockout/tagout to be integrated into a flexible


manufacturing system (FMS) control;
• use the model of Markov chain (homogeneous and non-homogeneous) by
incorporating different parameters, so the model will be as realistic as possible;
• demonstrate the complexity of flexible manufacturing system (FMS) by increasing
the number of machines along with number of products;
• present the resolution of a complex flexible manufacturing system (FMS) by the
resolution of HJB (Hamilton-Jacoby-bellman) differential equations and numerical
approach;
• take the subsequent conclusions.

Keyword: Flexible manufacturing systems; Production cost; Passive redundancy;


Lockout/tagout.
139

2. Assumptions

In this paper, we use assumptions as followed:

For one machine producing one type of part:

1- The costs of corrective maintenance of failure type 2 is more than corrective maintenance
cost of failure type 1.

2- The mean corrective maintenance time of failure type 1 is shorter than the mean corrective
maintenance time of failure type 2.

3- The machine, after each repairing becomes new, which means we consider the policy of as
good as new (AGAN).

4- Throughout this paper, we consider repairing (corrective maintenance of failure type 1 and
2) action with lockout/tagout.

For a system in passive redundancy (two machines non-identical) producing one type of part,
we consider previous assumptions and we add other assumptions as follows:

5- Machine N1 is the main machine, machine N2 is a standby machine.

6-The main machine, after each repairing (corrective maintenance of failure type 1 and 2)
returns to production immediately and the standby machine turns back at rest.

7- Both machines do not fall down at the same time.

8- The standby machine can be broken down through rest. (Exogenous Factors).

2.1. One machine producing one part type of a flexible manufacturing system (FMS):

For one machine producing one part type we refer to Charlot et al. (2006) studies. In this
paper we present only the homogeneous Markov chain model and a manufacturing system
consisting of two (2) machines (non-identical) in passive redundancy.

2.2. Two machines (non-identical) in passive redundancy producing one part type in the
flexible manufacturing system (FMS):
140

In this case, the dynamics of the system are the same as one machine producing one part
type, that is, discrete elements ζ (t ) and continuous elements x(t). The discrete elements ζ (t )
represents the machine’s state and continuous elements x(t) represent the stock level. Hence,
we have ζ (t ) ∈ M = {1, 2,3, 4,5} . A positive value of x(t) is inventory. A negative value is

backlog. It is assumed that the system meets a constant demand rate. In this part, two
machines with different production rates were considered as well as different failure rates.
(λ1 ≠ λ2 and U max
1 2
≠ U max )

Therefore, the transition diagram obtained (Figure 1) is as followed:

Figure 1 State transition diagram (Two machines (non-identical) in passive redundancy


producing one part type)
Table1 Presents the transition modes for figure 1.

ζ 1 (t ) 2 2 2 3 4
ζ 2 (t ) 1 3 4 2 2
ζ (t ) 1 2 3 4 5

In this paper, the machine can break down randomly.


141

The 5 X 5 transition matrix Q of our system is as followed:

q q12 q13 q14 q15 


 11 
 q 21 q 22 q 24 q 25 
 0 
Q =  q 31 q 33 q 34 q 34 , (1)
0
 
 k 41 q 42 q 43 q 44
0 
 
 k 51 q 52 q 53
0
q 55 

Hence, the transition matrix Q depends on:

k 41 : Corrective maintenance rate with lockout/tagout of failure type 1 of main machine;

k51 : Corrective maintenance rate with lockout/tagout of failure type 2 of main machine;

Where k41 = q 41 and k51 = q 51 .

Hence, we have Π (.) = (π 1 , π 2 , π 3 , π 4 , π 5 ) as representing the vector of limiting probabilities

from modes 1 to 5.

Submitted to the following:

qαβ (k 41 , k 51 ) ≥ 0 (α ≠ β ), (2)

q αα (k41 , k 51 ) = −  qαβ (k 41 , k 51 ) , (α , β ) ∈ M (3)


α ≠β
142

The set of acceptable decisions at mode α (t ) are defined by:

Γ (α ) = {( u1 (.), u 2 (.), k
4
41
(.), k 51
(.)) ∈ R ,
max max
0 ≤ u 1 (.) ≤ u 1 , 0 ≤ u 2 (.) ≤ u 2 ,
min
k 41 ≤ k 41 ≤ k 41 , k 51
max min
≤ k 51 ≤ k 51
max
}, (4)

The control policies (control variables) at mode α (t ) are u1 (.), u2 (.), k 41 (.) and k 51 (.) ;

The function cost is as followed:

{ ∞ −ρ t
J (α , x, u1 , u2 , k 41 , k 51 ) = E  e
0
g ( x, u1 , u2 , k 41 , k 51 ) dt x ( 0) = x, ξ ( 0) = α , }
∀u1 (⋅) & u2 ( ⋅) ∈ Γ (α ) , (5)

Hence, we have Π (.) = (π 1 , π 2 , π 3 , π 4 , π 5 ) as representing the vector of limiting probabilities

from modes 1 to 5.

The subject of equations (1)-(5) is minimizing J(.) given by equations (5) as followed:

ν ( x,α) =
u(α)∈Γ(α)
(
inf J α, x,u , u , k , k ,
1 2 41 51
) ∀α ∈M (6)

The main aim of this paper is to optimize the production cost, maintenance cost as well as
lockout/tagout cost. This policy is satisfied by the set of Hamilton-Jacobi-Bellman (HJB)
equations which are shown in Kenné et al. (2003), Charlot et al. (2006). We write the HJB
equation as followed:

 ∂ 
ρν ( x,α) = min ( u −d) ν ( x,α) + g ( x,u,α) +  qαβν ( x,β ), ∀α, β ∈M (7)
u∈Γ(α)  ∂x α≠β 

The optimal control policy (u , u , k


*
1
*
2
*
41
*
, k51 ) denotes the minimizer over Γ(α ) of the right
hand side of equations (5).
143

3. Numerical approach

In this section, we expand the numercial method for solving the optimality conditions
presented in the previous section. This method is based on the Kushner approach (Kushner
and Dupuis -1992), Boukas and Haurie (1990), Kenne et al. (2003).

In this paper , we used Kushner’s approach for an approximation scheme of the gradient of
value function ν ( x,α ) . Let h denote the length of the finite difference interval of the

variable x. Hence, using h, ν ( x,α ) is approximated by ν h ( x,α ) and ν x ( x,α ) as followed:

(
1 h h
) 
 h ν ( x + h, α ) −ν ( x, α ) × (u − d ) if ( u − d ) ≥ 0, 
ν x ( x, α ) × (u − d ) =  , (8)
( 1
 h
h h
)
 ν ( x, α ) −ν ( x - h, α ) × (u − d ) otherwise, 


After approximation given by equation (8) and some manipulations, the HJB equations (7)
can be rewritten as:

 −1
 u − d  
j   n uj −d j h
(
h n j
ν ( x,α ) = min  ρ + qαα +   ⋅  ν ( x + h,α ) K +
u∈Γh(α )  j=1 h j   j=1 h j
  

)
+ν h ( x − h ,α ) K − + g ( x, u,α ) + Q.ν h ( x ,..., xn ,α )
1 }

(9)

Where Γ (α ) is the discrete feasible control space or the so-called control grid and the other

term used in equation (9) is defined as:

1 if (u − d ) ≥ 0,
K+ = 
0 otherwise,

1 if (u − d ) < 0,
K− = 
0 otherwise.
144

The system of equation’s (9) can be interpreted as the infinite horizon dynamic programming
equation of a discrete-time, discrete-state decision process as in Kenné et al. (2003), Charlot
et al. (2006), for optimization of production and maintenance control problems. The next

theorem shows that the value function ν h ( x, α ) is an approximation of ν ( x, α ) for a small

step size h.

4. Numerical example and sensitivity analysis

We consider a manufacturing system for two (2) machines (non-identical) in passive


redundancy case with five (5) states, describing by the homogeneous Markove process
α ∈ M = [1, 2,3, 4,5] .The discrete dynamic programming equation (9) gives the equations
(10)-(14) for two (2) machines (non-identical) in the passive redundancy case producing one
type part:

 u −d 
 1 (ν h ( x+h,1) k + +ν h ( x-h,1) k − )
−1  
 u −d   h 
h
ν ( x,1) = min  ρ + 1 h h
+ q + q + q + q  + g ( x,1) + q ν ( x,2) + q ν ( x,3)  , (10)
u∈Γ(1)  h 12 13 14 15  12 13
   
+q ν h ( x,4) + q ν h ( x,5) 
 14 15 
 
 
 u1 − d h x+h,2 k + + ν h x-h,2 k − ) 
 
−1  h (ν ( ) ( )
u −d  
ν h ( x,2 ) = min  ρ + 1

+ q + q + q   + g ( x,2 ) + q ν h ( x,1) + q ν h ( x,4 )  , (11)
 21 24 25   21 24
u∈Γ  2   h 
 
   + q ν h ( x,5 ) 
 25 
 
 

 
u −d 
−1  1 (ν h x+h,3
( ) k + + ν h ( x-h,3 ) k − ) 

u −d
  h 
   
ν h ( x,3 ) = min ρ + 1 +q +q + q   + g x,3 + q ν h
( ) ( x,1) + q34ν h ( x,4 )  , (12)
u∈Γ  3   h 31 34 35   31 
   + q ν h x,5
 
 
 35 ( ) 

 
 
145

 
 u2 − d h + h − 

( )
−1  h (ν x+h,4 k +ν x-h,4 k )
 
( )
u −d 
ν h ( x,4) = min  ρ + 2 + k + q +
 
q   g ( x,4) q ν ( x,1) q ν ( x,2)  ,
+ + h + h (13)
 41 42 43   41 42
u∈Γ 4  
  h 
 
  +q ν h ( x,3) 
 43 
 
 

 
 u2 − d 
−1  (ν h x+h,5 k + +ν h x-h,5 k − ) 
( ) ( )
   h 
 u −d   
ν h ( x,5) = min 2 + k + q + q  + g x,5 + q ν h x,1 + q ν h x,2  ,
ρ +
 h 51 52 53   ( ) 51 ( ) 52 ( ) (14)
u∈Γ 5     
h x,3
 
  +
 53
q ν ( ) 

 
 

Here, we present results which were obtained through the different figures.

Figure 2 Average cost/backlog cost


146

In Figure 2, we plotted the average cost variation according to the backlog cost for one
machine and two (2) machines non-identical in passive redundancy case producing one part
type. In this figure the lockout/tagout, corrective maintenance rate of failure type 1 and type
2 are set to minimum and maximum values. Firstly, we note that the backlog cost variations
don’t affect much the average cost if we increase the lockout/tagout, corrective maintenance
rate for failure type 1 and type 2. Secondly, we observe, if we add the standby machine to our
system, we can reduce increasingly this average cost variation. As one can observe from
Figure 1, we are able to better control this variation with the maximal lockout/tagout,
corrective maintenance rate for failure type 1 and type 2.

Figure 3 Average cost/holding cost

In Figure 3, we want to verify the influence of holding cost increase according to the average
cost. Therefore as shows in the above figure, we can better control this evolution in passive
redundancy case where the rates of lockout/tagout, corrective maintenance for failure type 1
and 2 are set to their maximum values.
147

Figure 4 Average cost/corrective maintenance cost for failure type 2

In Figure 4, we plotted corrective maintenance cost of failure type 2 for the two (2) cases
under study. We do not have a significant variation of the average cost by increasing the
corrective maintenance cost of failure type 2, but we observe that the average cost is much
lower in the passive redundancy case compared to the other cases where the rates of
lockout/tagout, corrective maintenance for failure type 1 and type 2 are set to their maximum
values. We can argue as noticed in the preceding analysis (Figure 4), by increasing the
corrective maintenance cost for failure type 1, we observe that the average cost of passive
redundancy case is much lower than the other cases where the rates of lockout/tagout,
corrective maintenance for failure type 1 and 2 are set to their maximum values.
148

5. Conclusion

In this paper we verified the influence of lockout/tagout and corrective maintenance rate of
two types of failures for a passive redundancy case consisting of two (2) non-identical
machines producing one part type. We observed that the passive redundancy case allows us
to optimize in a better way the production cost and the maintenance cost while guaranteeing
occupational safety. This optimization is most useful, if we introduce an effective planning of
lockout/tagout and maintenance during the production control. In this paper, we considered
the hedging point structure for the control policy of our systems. We developed an effective
solution approach to determine the optimal production cost and corrective maintenance cost
including the lockout/tagout control. The model proposed in this manuscript might be used
for various industrial sectors. Further researches will verify the influence of lockout/tagout
and corrective maintenance rate for a production line consisting of two (2) machines
flowshop with internal buffers and one standby machine.

Acknowledgments

The work described in this paper was supported by the following funding agencies: RRSSTQ
(Réseau de recherche en santé et en sécurité du travail du Québec), FRSQ (Fonds de la
recherche en santé – Québec), FQRSC (Fonds québécois de la recherche sur la société et la
culture), FQRNT (Fonds de recherche sur la nature et les technologies – Québec) and
l’IRSST (Institut de recherche Robert-Sauvé en santé et en sécurité du travail).
149

References

Boukas, E.K. and Haurie, A. 1990, Manufacturing Flow Control and Preventive
Maintenance: A Stochastic Control Aproach, IEEE Transactions on Automatic
Control 33, pp. 1024-1031.
Charlot, E., Kenné, J.P. and Nadeau, S. 2006, Optimal Production, Maintenance and
Lockout/tagout Control Policies in Manufacturing Systems, Int. J. Production
Economics 107, pp. 435-450.
Gharbi, A. and Kenné, J.P. 2003, Optimal Production Control Problem in Stochastic
Multiple-Product Multiple-Machine Manufacturing Systems, IIE Transactions 35,
pp. 941-952.
Kushner, H.J. and Dupuis, P.G. 1992, Numerical Methods for Stochastic Control Problems
in Continuous Time, Springer, New York.
Kenné, J.P, Boukas, E.K. and Gharbi, A. 2003, Control of Production and Corrective
Maintenance Rates in a Multiple-Machine, Multiple-product Manufacturing
System, Mathematical and Computer Modelling 38, pp. 351-365.
ANNEXE 2

LOCKOUT/TAGOUT AND OPERATIONAL RISKS IN THE PRODUCTION


CONTROL OF A TRANSFER LINE WITH PASSIVE REDUNDANCY

Behnam Emami-Mehrgani1, Sylvie Nadeau1, Jean-Pierre Kenné1

1
Départment of Mechanical Engineering, École de technologie supérieure, Montreal
(Quebec) H3C 1K3, Canada

American Society of Mechanical Engineers (ASME)


November 11-17, 2011, Denver, Colorado, USA

Abstract

This paper presents an analytical model for the joint determination of optimal production and
occupational safety for a failure prone manufacturing system consisting of three machines
(two machines as passive redundancy and a third machine in series with the previous ones)
producing one type of part. these machines are subject to breakdowns and repairs and the
control problem is subject to non-negative constraints on work-in-processes (wip). the
decision variables are the production rate of two main machines and a standby machine. the
decision variables influence the wip levels, the inventory levels and the system’s capacity. the
system capacity is assumed to be described by a finite state markov chain. the aim of this
paper is to minimize the cost of wip, inventory while respecting occupational safety. the
proposed approach is based on the combination of analytical formalism, simulation
modeling, design of experiments and response surface methodology to optimize a transfer
line in passive redundancy producing one part type. the usefulness of the proposed approach
is illustrated through a numerical example.

Keywords: Production Control; Lockout/tagout; Passive redundancy; Simulation,


Experimental design; Response surface methodology
152

1. Introduction

The problem of optimally controlling the production rates of manufacturing system has been
widely discussed in the scientific literature. In the literature, several approaches, mainly
heuristic and optimal procedures, are employed to solve the problem. The first approach is
present in the research literature, recent surveys is addressed by Gupta and Stafford (2006).
On the other hand, the second approach, which consists of a stochastic optimal control
problem formulation, to determine optimal control policies for a given problem. The optimal
control problems of stochastic flow-shops with limited buffers producing one part type
considered as the theory foundation of the optimisation problem. In this context, Presman et
al. (1995), considered a production planning problem in an N-machine flow-shop subject to
breakdown and repair of machines and to nonnegativity constraints on work-in-process. The
authors show that the policy obtained minimizes the expected discounted cost of production
and inventory/backlog over an infinite horizon.

The same problem was considered in Sethi et al. (2000), to minimise the long-run average
cost. The authors used stochastic dynamic programming formulation and showed that the
value function of the problem is locally Lipchitz and the value function is a solution to a
dynamic programming equation jointly with a certain boundary condition. In this horizon,
Yavuz and Tufekci (2006), studied a real case study of an electronic manufacturing flow-
shop. The authors divided the master problem into two sub-problems in order to determine
the batch sizes and production sequences. They developed a dynamic programming procedure
in order to solve the batching problem and recommended an existing method to solve the
sequencing problem. They illustrated the efficiency of their approach to answer the just-in-
time (JIT) goals.

Manufacturing systems operate in a stochastic environment because the machines are subject
to random breakdowns and repairs. It is possible to predict and control certain events while
others occur randomly and are beyond control within manufacturing systems (Gershwin,
2002). A rising number of breakdowns and repairs degrade the dynamic of the production
systems, reduces their availability, and increase occupational hazards associated with
153

maintenance activities. Left unattended, these disruptive elements erode competitiveness


expressed in terms of quantity of products, quality of products and occupational safety.
Occupational safety researchers worldwide have carried out studies confirming the
importance of monitoring and controlling undesirable incidents or accidents during
maintenance procedures. An important consideration arises: How to optimize production yet
ensure occupational safety. One possible answer is lockout/tagout. It consists in locking a
machine with a padlock to discharge all sources of residual energy (hydraulic, electrical, etc.)
in order to avoid the premature starting of equipment throughout an intervention. Many
managers wrongly assume it takes too much time to plan and carry out a lockout/tagout,
fearing such downtime will reduce the productivity or the performance.

In this horizon, Emami-Mehrgani et al. (2011), considered an analytical model combining the
lockout/tagout, the production and corrective maintenance policies for a passive redundancy
system, consisting of two non-identical machines. Their work demonstrates clearly that
passive redundancy optimizes production and maintenance costs while enhancing
occupational safety. Even greater benefits occur if effective lockout/tagout maintenance
planning occurs in concert with production control. The main contributions of this work are
reduced the production cost and have free space-time to minimize the possibility of
circumvention of protection devices or retraction of lockout/tagout procedures for the
machines that are under repair.

In this paper, we resort to a numerical approach to find an approximate value function,


instead of the true value function, to construct the control policy. To illustrate the usefulness
of the proposed model, we show a numerical example for a transfer line with three non
identical machines (two machines as passive redundancy and one machine in series with the
previous ones). Based on the results given by numerical methods, a structure of the control
production policy is defined and parameterised herein by parameters called control
parameters. The proposed control approach consists of estimating the relationship between
the incurred cost and the control policy parameters considered as control factors. The
extension of hedging point structure, parameterized by these factors. The simulation model is
154

used to determine the related output (cost incurred), for each configuration of input factor
values. The significant effects of input factors are determined by experimental design and the
surface methodology is applied to the input-output data obtained in order to estimate the cost
function and the related optimum.

2. Assumptions and notations

This paper incorporates the following assumptions and notations:

2.1 Assumptions

1. The corrective maintenance is carried out with lockout/tagout.

2. The main machine is more robust than the standby machine.

3. The main machine and the standby machine produce the same type of parts for the work-
in-process (WIP).

4. The main machine returns to production immediately after each repair (corrective
maintenance with lockout/tagout) and the standby machine stands idle.

Assumption 4 is a classical assumption in passive redundancy system. It is due to the nature


of a passive redundancy system.

2.2 Notations

The following notations are used in the rest of this article:

x1 (⋅) inventory levels of work-in-process

x2 (⋅) inventory/backlog levels of finished products


+
c1 holding cost per unit of item over per unit of time for work-in-process
+
c2 holding cost per unit of item over per unit of time for finished products

c2 backlog cost per unit of item over per unit of time for work-in-process
155

α
c cost incurred for the operation on the machine under repair at mode α
cr1 corrective maintenance cost of main machine 1

cr2 corrective maintenance cost of main machine 2

cr3 corrective maintenance cost of standby machine

c lockout/tagout cost
tagout

u r1 corrective maintenance rate with lockout/tagout of main machine 1

u r2 corrective maintenance rate with lockout/tagout of main machine 2

urs maintenance rate with lockout/tagout of standby machine

g (⋅) instantaneous cost


J (⋅) total cost
v (⋅) value function
ρ discount rate
d demand rate
ui (⋅) production rate of the machine i ( i =1, 2, s )

ui
max
(⋅) maximal production rate of the machine i ( i =1, 2, s )

1,2
q12 main machines M1 and M 2 failure rate

s
q12 standby machine M s failure rate

3. Problem statement

In this paper we consider the flow control problem for a tandem production system with
passive redundancy ( M = 3) . The system is shown in Figure 1. The main machines have two
states: ξ1,2 (t ) = 1 if the main machine M i ( i = 1, 2) is operational and ξ1,2 (t ) = 2 if the main

machine M i ( i = 1, 2) is under repair. The standby machine has three states: ξ s (t ) = 1 if the

standby machine M s is operational, ξ s (t ) = 2 if the standby machine M s is under repair and


156

ξ s (t ) = 3 if the standby machine M s is at time-off. We use ui (t ) to denote the input rate to

M i and x1 (t ) to denote the number of parts in the buffer between M i and M 2 (i = 1, s ) . Note that the

control problem is subject to non-negative constraints on work-in-process (WIP) that means


+
x1 (t ) ≥ 0. For buffer x1 (t ) the difference is always positive (i.e. inventory costs c1 are thus

charged) or equal to zero (i.e. starvation of machine M 2 ). The difference between actual

production and downstream demand at any time represents by x2 (t ) . For buffer x2 (t ) the
difference between actual production and downstream demand is positive (i.e. inventory
costs c2+ are thus charged) or negative (i.e. backlog costs c2− are thus charged).

Figure 1 Production line with passive redundancy one part-type producing

Figure 2 State transition diagram


157

Figure 2 displays the modes of the system associated to the process ξ (t ) .

The dynamics of the system can be written as follows:

x ( t ) = u ( t ) − u ( t ) = u ( t ) Ια + u s ( t ) Ια
s − u 2 ( t ), x (0) = x , (1)
1 1 2 1 1 1 1
with :
 1 if α = 3, 4, 5, 6, 7, 8,  1 if α = 1, 2,
Ια =  and Ια s =
1 0 otherwise,  0 otherwise.

x2(t) =u2(t)−u3(t) =u2(t)−u3(t), x2(0) =x2, (2)


with:
u3(t):=d.

In matrix notation, the system of equation (1)-(2) becomes:

x ( t ) = A u ( t ) , x (0 ) = x, (3)

1 −1 0 
w h e re A =  , u ( t ) = ( u ( t ) , u ( t ) , u ( t )) = ( u ( t ) Ι α + u s ( t ) Ι α
s , u 2 ( t ), u 3 ( t ))
0 1 − 1  1 2 3 1 1
a n d x ( t ) = ( x ( t ), x ( t ) ),
1 2
w ith :
1 if α = 3, 4 , 5 , 6 , 7 , 8 , 1 if α = 1, 2 ,
Ια =  , Ια
s =  and u ( t ) := d .
1 0 o th e rw is e , 0 o th e r w is e , 3

The transition rates matrix of the stochastic processes ξ (t ) are denoted by Q such that

Q= q { αβ } , with qαβ > 0 if α ≠ β and qαα =  β ≠α qαβ , where α , β ∈ Β.


The transition probabilities associated to qαβ are expressed as:

qαβ (⋅)δt+0(δt) if α≠β,



p ξ(t +δt) = β ξ(t) =α =  (4)
1+q (⋅)δt+0(δt) if α=β.
 αβ

The transitions rates matrix Q is expressed as follows:


158

 q1 1 q
12
0 0 q
15
0 q
17
0 
 
q21 q
22
0 0 0 q
26
0 q
28 
 
 0 0 q
33
q
34
q
35
0 q
37
0 
 0 0 q q 0 q 0 q 
 43 44 46 48 
q , (5 )
0 q 0 q5 5 q 0 0 
 51 53 56 
 0 q 0 q q q 0 0 
 62 64 65 66 
q 0 q 0 0 0 q77 q 
 71 73 78 
 0 q
82
0 q
84
0 0 q
87
q
8 8 

The operational mode of the whole system can be described by the random vector
ξ (t ) = (ξ1 (t ), ξ 2 (t ), ξ s (t )) taking values in Β = {1, 2,3, 4,5, 6, 7,8} .
Without loss of generality, for the three machines flow-shop case in passive redundancy ξ (t )
can be expressed as follows:

1 M isunderrepair, M isoperationaland Ms isoperational;


1 2

2 M is under repair, M isunderrepairand Ms isoperational;
1 2

3 M is operational, M is operational and Ms is underrepair;
1 2
4 M is operational, M is under repair and Ms isunderrepair;
 1 2
ξ (t) = 
5 M is under repair, M is operational and Ms is underrepair;
1 2
6 M is under repair, M is underrepairand Ms isunderrepair;
 1 2
7 M is operational, M is operational and Ms is at time-off;
 1 2
8 M is operational, M is under repair and Ms is at time-off.
 1 2

The set of admissible decisions at mode α (t ) and control policies (control variables) at mode
α (t ) :

 ((u1 (⋅), u 2 (⋅), u s (⋅)) ∈ R 3 , 


 
Γ (α ) =  0 ≤ u1 (⋅) ≤ u1max , 0 ≤ u 2 (⋅) ≤ u 2max ,  , (6)
 0 ≤ u (⋅) ≤ u max 
 s s 

In equation (6), u1max is the maximal production rate of the main machine 1, u2max is the

maximal production rate of the main machine 2 and u smax is the maximal production rate of
159

the standby machine. Γ(α ) denote the set of all admissible controls with respect to x ∈ S

and α (0) = α . Let S = [ 0, ∞ ) × R ⊂ R m denote the state constraint domain. The control

problem consists of finding an admissible control law u (⋅) = (u1 , u2 , us ) that minimize the cost

function J (⋅) given by:

∞ −ρ t 
J (α, x, u) = E   e g (α, x, ⋅) dt x ( 0) = x, ξ ( 0) = α , (7)
0 

g(x, α , ⋅) = c1+ x1+ + c2+ x2+ + c2− x2− + cα

Where is the instantaneous cost, c + , c − and cα , being the cost per unit to produce parts
for inventory, backlog as well as intervention cost on the machine.

x+ = max {0, x} , x− = max {− x,0} and cα = ((cr1 + ctagout )ur1 )Ind {α = 1} + ((cr1 + ctagout )ur1
+(cr2 + ctagout )ur2 )Ind {α = 2} + ((crs + ctagout )urs )Ind {α = 3}
+((crs + ctagout )urs + (cr2 + ctagout )ur2 )Ind {α = 4}

+ (( c rs + ctagout ) u rs + ( c r1 + ctagout ) u r1 )Ind {α = 5}


+ (( c rs + ctagout ) u rs + ( c r1 + ctagout ) u r1
+ ( c r2 + ctagout ) u r2 )Ind {α = 6}
+ (( c r2 + ctagout ) u r2 )Ind {α = 8}
with
1 if Θ (⋅) is true 
Ind {Θ (⋅)} =  
0 otherwise 

Letν ( x,α ) denote the value function or minimum discounted cost for equations (7) as
expressed in the following equation:

ν ( x,α ) = min J (α , x, u ) , ∀α ∈ Β (8)


u∈Γ( x,α )

The value function v(⋅) given by (8) satisfies a set of partial differential equation known as
the Hamilton-Jacobi-Bellman (HJB) equations in terms of directional derivatives (DD).
160

4. Optimal conditions: dynamic programming equation

A necessary and sufficient condition characterizing the optimal control policy is described by
the set of partial differential equation known as the Hamilton-Jacobi-Bellman (HJB)
equations in terms of directional derivatives (DD), as in Presman et al. (1995), by:

 
ρν ( x,α ) = min v'  ( x, α ) + g (α , x, ⋅) +  q ν ( x,β ) −ν ( x,α )   , (9)
u∈Γ( x,α )  Au α ≠ β αβ 
∀α , β ∈ Β

Where:
v'  ( x, α ) = (u1 − u2 )v x1 ( x, α ) + (u2 − u3 ) × v x2 ( x, α ).
Au

Note that v x1 and v x2 are the partial derivatives of v compared to x1 and x2 respectively. The

optimal control policy (u1* , u2* , us* ) denotes a minimizer over Γ (α ) of the right-hand side of
equation (9). This policy corresponds to the value function described by Equation (8). Then,
when the value function is available, an optimal control policy can be obtained as in Equation
(9). However, an analytical solution of Equation (9) is almost impossible to obtain.

5. Numerical approach

In this section, we use the numerical method for solving the optimality conditions presented
in the previous section. This method is based on the Kushner approach (Kushner and
Dupuis,1992) .The basic idea behind this approach consists of using an approximation
scheme for the directional derivative of the value function v ( x , α ). Let h1 and h2 denote the

length of the finite difference interval of the variables x1 and x2 . Hence, using h1 and h2 ,

v ( x , α ) is approximated by v ( x, α ) , v x1 and v x2 are approximated by:


h
161

(
1 h h
) 
 h ν ( x1, x1 +h1,α) −ν ( x1,α) ×(u1 −u2) if ( u1 −u2 ) ≥0 
vx1()⋅ =vu' −u ()⋅ = 1 , (10)
1 2

(
1 h h
)
 h ν ( x1,α) −ν ( x1, x1-h1,α) ×(u1 −u2) otherwise
1



with:
1 if α =3,4,5,6,7,8 α 1 if α =1,2
u1 =u1Ι1α +usΙαs , Ι1α = , Ιs = and u2 =u2.
0 otherwise 0 otherwise

1 h
( h
)
h ν ( x1, x2 +h2,α) −ν ( x1,α) ×(u2 −u3) if (u2 −u3) ≥0


vx2 ()⋅ =vu' −u ()⋅ = 2 , (11)
2 3 1 h
( h
)
h ν ( x1,α) −ν ( x1, x2-h2,α) ×(u2 −u3) otherwise
2



with:
u2 =u2 and u3 =u3 :=d.
We manipulated the approximation arrived at in equations (10) and (11) to rewrite the
HJBDD equation (9) as follows:
  
  
   q (vh ( x, β )) + g ( x,.) 
  β ≠α αβ 
  

 (u − u ) (u − u
 )
ν h ( x, α ) = min  ρ + qαα + 1 2 + 2 3   1 2 

−1 
(u − u
 ) 1 1 2 (
(ν h x + h , x , α k +   
1   )
, (12)
 −  
h 
u∈Γ ( x,α ) 

h
1
h
2

 
h h
(
1  +ν x1 - h1, x2 , α k1 )    )
 

 + 2 3 
 h
 (u − u ) (ν x , x + h , α k
1 2 2
+ 
(
2  
 )
  h  




h
2 
 1 2 2 (
+ν x , x − h , α k − )   
2   )
Where Γ h (α ) is the discreet feasible control space and the other terms used in equation (12)

are defined as:

1 if (u1 − u2 ) ≥ 0, 1 if (u1 − u2 ) < 0,


K1+ =  and K1− = 
0 otherwise, 0 otherwise,

1 if (u2 − u3 ) ≥ 0, 1 if (u2 − u3 ) < 0,


K2+ =  and K2− = 
0 otherwise, 0 otherwise,
with:
u2 = u2 and u3 = u3 := d .
162

In this paper, we use the policy improvement technique to derive a solution of the
approximating optimization problem. The algorithm of this technique can be found in
Kushner and Dupuis, (1992).

6. Numerical example and sensitivity analysis

In this section, we consider a transfer line with three non identical machines (two machines
as passive redundancy and one machine in series with the previous ones). The system
capacity is described by an eight Markov process with states ξ (t ) ∈Β = 1,2,3,4,5,6,7,8 .The
discreet dynamic programming equation for our manufacturing system calculated by
equation (12), with α = 1,2,3,4,5,6,7,8 . We use the following computational domain:

Gxh = {( x1 , x2 ) : 0 ≤ x1 ≤ 5; − 5 ≤ x2 ≤ 5} ,

The parameters for our case study appear in Table 1.

Table 1. Parameters of the numerical example

Parameter c1+ c2+ c2− 1,2


q12 q12s u r1
Value 1 10 150 0.02 0.03 0.08
Parameter ur2 u rs max
u
1 u max
2 u max
s
ctagout
Value 0.08 0.08 1.3 1.25 1.2 50
Parameter cr1 cr2 crs ρ d
Value 150 200 250 0.2 1

The results obtained for the set of control variables u1 , u2 and us of a production line with
passive redundancy system are given in Figure 3 for an illustration purpose where M1, M2 are
operational and M3 (standby machine) is under repair.
163

(a) Production rate of M1

(b) Production rate of M 2

Figure 3 Production rate of M 1 and M 2 at mode 3.

Figure 3 shows that there is no need to produce the part with sufficient stock levels both in the
work-in-process (WIP), described by x1 and in the final stock described by x2. For small stock
levels, the obtained policy defines well the region in the domain (x1,x2) where a maximal
164

production rate is optimal. These results allow us to find our policy. As illustrated in Figure 3,
we have:

 uimax if x1 < Z1 & x2 < Z 2 


  
ui ( x1 , x2 ,3) =  d if x1 = Z1 & x2 = Z 2 
  0 otherwise 
   ,
  (13)
 u2max if x2 < Z3 & x1 > 0  
  
 u1 ( x1 , x2 ,3) =  d if x2 = Z 3 & x1 > 0  
  0 otherwise 
  
With i = 1, s

Note that our policy to find the optimal value of production rate ui with i = 1, 3 depends on two

factors Z1 and Z2 , and the optimal value of production rate u2 depends on one factor such Z3 .
The next sections are aimed at developing a systematic approach for determining optimal
values of Z1 , Z2 , Z3 .

7. Control approach, experimental design and response surface methodology

In the sphere control theory, results from traditional planning methods of production of
flexible manufacturing system (FMS) are not sufficient to reach a comfortable level of
desired performance. To improve this method, the descriptive capacities of conventional
simulation model are combined to analytical models, experimental design and response
surface methodology. For more information, we refer the reader to Kenné and Gharbi,
(2001).
165

The structure of the proposed control approach is presented in Figure 4.

FMS Control Analytical Numerical


problem Model Methods

* * *
Z1 , Z 2 , Z3 Z1 , Z2 , Z3

Response Experimenta Simulation


Surface l Design model

Figure 4 Diagram of control approach

To illustrate the approach presented in this paper we consider a production line with passive
redundancy consisting of three machines, producing one part type. The simulation parameters
used in this paper are the same as the Table 1. For three factors problems, as illustrated in
previous sections, we selected a 33 response surface design since we have three independent
variables at three levels. This design leads to the completion of 27 experimental trials. We
have three independent variables Z1 , Z2 and Z3 where Z1 , Z2 and Z3 present our policy concerning

the stock level of x1 and x2 for the machine M1, M2, M3(standby) and one dependent variable
(the production average cost). The levels of the independent variables are presented in the
Table 2.

Table 2 Levels of independent variables

Factor Low level Medium level High level


Z1 5 15 25
Z2 5 15 25
Z3 5 15 25

In this paper, we chose three replications; hence we have 108 (27 × 4) simulation runs. For
more information, we refer the reader to Montgomery (2005). Thereafter, we considered all
possible combinations of different levels of independent variables by response surface
166

design. The objective of this design is to understand the effects of independent variables on
performance measures, in our case the production average cost. From the ANOVA table, the
independent variables Z1 , Z 2 , Z3 and interaction effect as well as the quadratic effect are
significant for the dependant variable at 0.05 level of significant. The R2 value of 0.88, that
means 88% of the total variability, is explained by the model.

The estimation of the regression coefficients is performed and the ten values achieved.

β0 = 860, β1 = 1,899, β2 = −18, β3 = 4,080, β11 = 0,174, β12 = −0,086, β13 = −0,351,
β22 = 0,530, β23 = 0, 272, β33 = −0,153.

The average cost function is given by:

Average cost = β0 + β1Z1 + β2 Z2 + β 3 Z3 + β11Z12 + β12 Z1Z2 + β13 Z1Z3 + β22 Z22 + β23 Z 2 Z3 + β33Z32

The minimum average cost function is located at Z1* = 14, Z2* = 14, Z3* = 15 where Z1* , Z 2* and Z 3*

represent the optimal values of independent variables Z1 , Z 2 and Z3 .

Figure 5 illustrates the contour plot of the average cost function or response surface.

Figure 5 Contour plot of the response surface

These values determined the hedging point policy for our manufacturing system where the
average cost is minimised and this control policy is the best approximation of the optimal
167

control. Without loss of generality of this proposal, this model is based on certain
assumptions relating to a transfer line consisting of three machines which are operated in
passive redundancy. We observed that the passive redundancy case allows us to optimize in a
better ways transfer line’s production cost while guaranteeing occupational safety.
Integration of the second machine as the passive redundancy makes possible to answer to
demand permanently. In addition this integration will be released the essential space-time to
intervention on the machine under repair, in order to minimize the possibility of
circumvention of protection devices or retraction of lockout/tagout procedures.

8. Conclusion

This paper confirms that it is possible to integrate a passive redundancy system in the
production line in order to: 1) increase the productivity of workers and material resources, 2)
optimize in a better way the production costs while guaranteeing occupational safety. In this
paper, we showed that our policy is an extension of the hedging point policy. Analytical and
simulation models were developed to describe the dynamic production under the hedging
point policy. An experimental design was used to determine the effects of the independent
variables on the average cost over the production horizon. We combined the analytical,
simulation and statistical methods to provide the estimation of the average cost related to the
control problem. The estimation of the average cost permits to calculate the best values of
control parameters. Finally, we showed that the passive redundancy systems improve the
safety of workers and production cost jointly. Taking into account certain conditions, the
proposed approach could be adopted in several industrial sectors.

Acknowledgments

We gratefully acknowledge for financial support of this work from the following agencies:
Natural Sciences and Engineering Research Council of Canada (NSERC); Réseau de
Recherche en Santé et en Sécurité du Travail du Québec (RRSSTQ); Fonds de la Recherche
en Santé – Québec (FRSQ); Fonds Québécois de la Recherche sur la Société et la Culture
168

(FQRSC); Fonds Québécois de la Recherche sur la Nature et les Technologies (FQRNT) and
Institut de Recherche Robert-Sauvé en Santé et en Sécurité du Travail (IRSST).
169

References

Gupta, J.N.D and Stafford, E.F., 2006. "Flowshop scheduling research after five decades".
European Journal of Operational Research, 169, 699–711.
Presman, E., Sethi, S., and Zhang, Q., 1995. "Optimal feedback production planning in a
stochastic N-machine flowshop". Automatica, 31 (9), 1325–1332.
Sethi, S.P., Zhang, H., and Zhang, Q., 2000. "Hierarchical production control in a stochastic
N-machine flowshop with long-run average cost". Journal of Mathematical Analysis
and Applications, 251, 285–309.
Yavuz, M. and Tufekci, S., 2006. "Dynamic programming solution to the batching problem
in just-in-time flowshops". Computers & Industrial Engineering, 51, 416–432.
Gershwin, S.B., 2002. "Manufacturing Systems Engineering". Prentice-Hall, Englewood
Cliffs, NJ, pp.267.
Emami-Mehrgani, B., Nadeau, S and Kenné, J.P., 2011 "Lockout/tagout and operational risks
in the production control of manufacturing systems with passive redundancy". To
appear in International Journal of Production Economics.
Kushner, H.J. and Dupuis, P.G., 1992. "Numerical Methods for Stochastic Control Problems
in Continuous Time". Springer, New York.
Kenné, J.P. and Gharbi, A. 2001. "A simulation optimization approach in production
planning of failure prone manufacturing systems". Journal of Intelligent
Manufacturing, New York; Kluwer Academic Publishers.
Montgomery, D.C. 2005. "Design and Analysis of Experiments". sixth edition, Wiley, New
York, NY.
ANNEXE 3

LOCKOUT/TAGOUT AND HUMAN ERROR IN PRODUCTION CONTROL OF


MANUFACTURING SYSTEMS WITH PASSIVE REDUNDANCY

Behnam Emami-Mehrgani1, Sylvie Nadeau1, Jean-Pierre Kenné1


1
Départment of Mechanical Engineering, École de technologie supérieure, Montreal
(Quebec) H3C 1K3, Canada

Gesellschaft für Arbeitswissenschaft (GFA)


22-24 febraury,2012, Kassel,Germany

Abstract

Lockout/tagout consists in locking a machine with a padlock to discharge all sources of


residual energy (hydraulic, electrical, etc.) in order to block premature starting of equipments
during a maintenance intervention. This paper is different compared to other research
projects on lockout/tagout. This paper illustrates the human error influence on lockout/tagout
as well as on maintenance activities. The aim of our model is to minimize production,
inventory, backlog and maintenance costs over an infinite planning horizon. We have
integrated lockout/tagout in the context of a manufacturing system using a control theory
based on mathematical models. More precisely, we consider a manufacturing system
consisting of two non-identical machines in passive redundancy producing one type of part.
These machines are subject to breakdowns and repairs. The system capacity is assumed to be
described by a finite state Markov chain. The decision variables are the production rate and
preventive maintenance rate. The decision variables influence the inventory level and the
system’s capacity. In the proposed model, the failure rate of the manufacturing system
depends on its age that means the preventive maintenance policy is machine age-dependent.
A numerical example and sensitivity analysis are presented to illustrate the usefulness of the
proposed approach.
172

Keywords: Lockout/tagout, Human error, Production control, Preventive maintenance,


Passive redundancy.

1. Introduction

The current use of lockout /tagout has several shortcomings such as lack of validation of its
feasibility and profitability. There is always a risk during intervention on machines that are
down. The risk of human error during these interventions has a direct impact on the
availability of manufacturing systems and can lead to occupational incidents or accidents.
Protection devices are often absent or bypassed during these interventions, which explains
the statistics of accidents remain high in this sector (Chinniah & Champoux 2008).

In this horizon, Emami-Mehrgani et al. (2011) considered an analytical model combining


lockout/tagout, production and corrective maintenance policies for a passive redundancy
system, consisting of two non-identical machines. Their work demonstrates clearly that
passive redundancy optimizes production and maintenance costs while increasing the
security level. Another work integrating lockout/tagout into operational risk in production
control is proposed by Emami-Mehrgani et al. (2012). They considered a manufacturing
system consisting of three machines (two machines with passive redundancy, and one in
series with the previous ones) producing one part type. Their work confirms that it is possible
to integrate a passive redundancy system in a production line in order to reduce the
production cost and have free space-time to minimize the possibility of circumvention of
protection devices or retraction of lockout/tagout procedures for the machines that are under
repair. The main contributions of this work are reduced production cost, preventive
maintenance cost and also is verified the influence of human error during maintenance
activity on the optimal safety stock levels.

2. Problem statement

This section develops a manufacturing system for two machines that are not identical in
passive redundancy and producing one type of part. In the system under consideration, the
173

main machine operates any time. The other machine, named as the cold-standby redundancy,
is in standby position and put online immediately when the main machine fails. The main
machine has two states: ξ1 (t ) = 1 if the main machine M is operational and ξ1 (t ) = 2 if the

main machine M is under repair. The standby machine has five states: ξs (t ) = 1 if the

standby machine S is operational, ξ s (t ) = 2 if the standby machine S is under repair,

ξs (t ) = 3 if the standby machine S is under preventive maintenance without human error,


ξs (t ) = 4 if the standby machine S is under preventive maintenance with human error and
ξs (t ) = 5 if the standby machine S is at time-off. The dynamics of the system can be written
as follows:

dx
= ui (⋅) − d , x (0) = x, with i = 1, S
dt
da
= f (u (⋅)), a (0) = a, a (T ) = 0. (1)
dt

The operational mode of the whole system can be described by a random vector
ξ (t ) = (ξ1(t ),ξs (t )) taking values in Β = {1, 2,3, 4,5} . Without loss of generality, for the
passive redundancy, ξ (t ) can be expressed as follows:

1 M is operational and S is under repair;


2 M is operational and S is under preventive maintenance without human error;

ξ (t ) =  3 M is operational and S is under preventive maintenance with human error;
4 M is operational and S is at time-off;

5 M is under rapair and S is operational.

24 34
The control problem consists in finding an admissible control law u (⋅) = (u1 , us , w42 , w42 ) that

minimizes the cost function J (⋅) given by:

∞ −ρ t 
J ( a, x,α,u1,us, w4224, w4234 ) = E  e g( a, x,α,u1,us, w4224, w4234 ) dt x( 0) = x, ξ ( 0) = a, a(0) =0, (2)
0 
174

Letν ( a, x,α ) denote the value function or minimum discounted cost for equations (2) as
expressed in the following equation:
J ( a, x,α,u1,us, w42 , w42 ) , ∀α ∈Β
24 34
ν ( a, x,α) = min (3)
,w42)∈Γ( x,α)
24 34
(u1,us,w42
Properties of the value function v (⋅) given by equation (3) is presented in Boukas and Haurie
(1990). It is shown that the value function v (⋅) given by (3) should satisfy a set of partial
differential equations known as the Hamilton-Jacobi-Bellman (HJB) equations.

3. Numerical example and sensitivity analysis

Let us consider a manufacturing system with two non-identical machines in passive


redundancy. The system capacity is described by a five Markov process with states
ξ (t ) ∈ Β = [1,2,3,4,5] . The parameters for our case study appear in Table1.

Table 1 Parameters of the numerical example

u1max u smax 24min


w42 24max
w42 34min
w42 34 max
w42 q14 q23 q24 q34 q54 d ρ

.32 .28 10-6 .06 10-6 .06 .05 .02 .08 .13 .05 .25 .01

24 34
The results obtained for the control variables u1 , us , w42 and w42 of a passive redundancy
system are given in Figures (1)-(2) for illustration purposes.

Figure 1 Main machine production rate and standby machine preventive maintenance rate
without human error at mode 2
175

Figure 1 shows the existence of different zones for the production rate of the main machine.
According to this figure in the area where the machine age is not yet advanced the production
rate must be maintained at zero, despite a low inventory level. On the other side, if the
machine age exceeds its youth with the same inventory level, the policy is suggesting to
produce with the demand rate. Beyond a certain age, the threshold becomes more important.
The policy is suggesting to produce with maximum rate below this threshold. In this area the
age leads to frequent failure. Figure 1 also shows the standby machine preventive
maintenance rate without human error, plotted in this figure, divides the computational
domain (x,a) into two regions where the preventive rate is set to its maximal value for
backlog situation and to zero for large stock levels. For significant stock levels, the zone in
the domain (x,a) where the preventive maintenance is set to its maximal value increases with
the machine age.

Figure 2 Main machine production rate and standby machine preventive maintenance rate
with human error at mode 3

Figure 2 illustrates the same structure as Figure 1, but the policy suggests keeping more
inventories by increasing the machine age, because the preventive maintenance has been
done with human error.
176

4. Conclusions

This paper sought to verify the influence of human error on the preventive maintenance
activity with lockout/tagout. In this paper, the control policy is based on an extension of the
hedging point structure. This work demonstrates clearly that human error during maintenance
activities can increase the production cost while reducing the safety of workers. So to avoid
this problem, we must better train our maintenance technicians and also we must not increase
the production cadence brutally. Lastly, a number of conditions must be met to make
effective use of this model.
177

References

Boukas, E.K. and Haurie, A. 1990, Manufacturing Flow Control and Preventive
Maintenance: A Stochastic Control Aproach, IEEE Transactions on Automatic
Control 33, pp. 1024-1031.
Chinniah, Y. & Champoux, M. 2008, La sécurité des machines automatisées : Analyse des
risques et des moyens de protection sur une presse à injection de plastique, R-557.
Montréal: Institut de recherche Robert-Sauvé en santé et en sécurité du travail
(IRSST), pp. 93.
Gershwin, S.B., 2002, Manufacturing Systems Engineering. Prentice-Hall: Englewood Cliffs,
pp. 267.
Emami-Mehrgani, B., Kenné, J.P. & Nadeau, S. 2012, Lockout/tagout and Optimal
Production Control Policies in Failure-prone Non-homogeneous Transfer Lines
With Passive Redundancy, International Journal of Production Research, Accepted
(January 2011).
Emami-Mehrgani, B., Nadeau, S. & Kenné, J.P. 2011, Lockout/tagout and operational risks
in the production control of manufacturing systems with passive redundancy,
International Journal of Production Economics, 132, pp.165-173.
ANNEXE 4

CADENASSAGE ET BAISSE DES COÛTS DE PRODUCTION

Behnam Emami-Mehrgani, Sylvie Nadeau, Jean-Pierre Kenné


Département de génie mécanique, École de technologie supérieure,
Montréal (Québec) H3C 0J9
Travail et Santé,2011, 27(2), pp.13-15

Cet article vise à intégrer le contrôle du C/D dans la gestion de la capacité de production.
Cette intégration améliore la rentabilité de l’entreprise, tout en augmentant le niveau de
sécurité des divers travailleurs.

1. Quelques statistiques importantes

Afin d’obtenir une estimation globale de l’impact de la maintenance sur les risques pour les
travailleurs au Québec, une étude a été effectuée par l’Institut de Recherche Robert Sauvé en
Santé et Sécurité du Travail (IRSST) sur les accidents du travail mortels au Québec entre
1999 et 2003 (Giraud et al., 2008). Les rapports de la CSST révèlent la même période, 1275
décès suite aux accidents du travail. Ils précisent que 163 d’entre eux (13%) ont été produits
pendant une activité de maintenance. Depuis quelques années, pour résoudre ce problème,
certaines entreprises ont mis en place un système de contrôle des risques par le biais du C/D
(Charlot et al.,2006). Leur application réelle est parfois escamotée faute de temps, au nom de
la performance de l’entreprise et à cause de difficultés technologiques.

2. Cadenassage/décadenassage

La norme canadienne CSA Z460-05 (2005) définit le C/D ou consignation/déconsignation


(Europe) comme «l’installation d’un cadenas ou d’une étiquette sur un dispositif d’isolement
des sources d’énergie conformément à une procédure établie, indiquant que le dispositif
d’isolement des sources d’énergie ne doit pas être actionné avant le retrait du cadenas ou de
l’étiquette conformément à une procédure établie».
180

3. Démarrage prématuré

Les accidents liés aux machines coûtent cher en vies humaines, en arrêts maladies et en coûts
divers. L’importance de l’intégration du C/D lors des interventions sur les machines n’est
plus donc à démontrer. L’augmentation des précautions de sécurité réduit la fréquence des
accidents, mais, elles augmentent le coût total de production. Notre question de recherche est
de savoir comment trouver une solution adéquate pour optimiser la production, tout en
augmentant la sécurité des travailleurs dans une entreprise par le biais du C/D?

4. Environnement incertain

Notre recherche sur le C/D vise à intégrer le contrôle du C/D dans la gestion de la capacité de
production en utilisant une théorie de commande, basée sur des modèles mathématiques
sophistiqués, mais existants (Emami-Mehrgani et al., 2011). L’environnement des systèmes
manufacturiers est incertain. La nature incertaine de l’environnement manufacturier découle
des machines sujettes à des pannes et des réparations aléatoires. Ces dernières engendrent des
accidents difficilement prévisibles lors des interventions de maintenance. Le projet vise à
intégrer le concept novateur promettant des avancées en gestion des activités de C/D dans la
gestion des capacités. Ce concept consiste à intégrer un système en redondance passive dans
le contexte manufacturier afin de mieux gérer les activités de C/D (Emami-Mehrgani et al.,
2010).

5. L’influence d’un système en redondance passive

D’abord, nous exposons l’influence du temps de C/D pour un système manufacturier


constitué d’une machine produisant un type de pièce. Ensuite nous assimilons le C/D pour un
système en redondance passive produisant un type de pièce. La redondance est qualifiée de
passive quand les éléments abondants ne sont mis en service qu'au moment du besoin; cela
signifie que parmi un ensemble d’éléments, seul, un sous-ensemble de ces éléments est en
service. Ceci implique que certains éléments seront en réserve ou en stock (Basil et
Dehombreux, 2002). Finalement, nous faisons une comparaison entre les résultats des deux
181

systèmes étudiés. Pour une machine produisant un type de pièce, quand la machine est
opérationnelle, elle peut être en mode production ou arrêt.

Nous supposons que notre système tombe en panne selon deux types de défaillances, l’une se
produit avec une fréquence moins élevée que l’autre. Quand la machine est en panne, pour
assurer la disponibilité de machine, on la répare. Nous admettons que le C/D se fait pour
effectuer chaque intervention de maintenance sur la machine. Dans le deuxième cas, nous
prenons en considération un système de production en redondance passive constitué de deux
machines non-identiques. Cette fois-ci, les machines tombent en panne selon deux types de
défaillances et le C/D est inclus dans la réparation. Pour établir les données, nous considérons
que le temps moyen de C/D varie entre un minimum et un maximum en fonction du niveau
de stock. Dans la réalité de ces systèmes, il en est tout autrement : le temps moyen de C/D est
considéré constant. La politique de commande obtenue est basée sur la politique à seuil
critique telle que définie dans (Akella et Kumar,1989). Cette politique recommande de
produire au taux maximal si le niveau d’inventaire est inférieur au seuil critique (3.5 sur la
Figure1); produire au taux minimal si le niveau d’inventaire est supérieur au seuil critique;
produire au taux de la demande si le niveau d’inventaire est égal au seuil critique.

Figure 1 Politique à seuil critique


182

Premièrement, nous montrons que pour un système manufacturier constitué d’une machine
produisant un type de pièce (M1P1), le C/D avec un temps minimal, comparé à un temps
maximal, a pour conséquence la diminution du niveau d’inventaire (Figure 2), et diminution
du coût de production (Figure 3).

Figure 2 Niveau d'inventaire/coût de pénurie

Figure 3 Coût moyen de production/coût de pénurie


183

Cette stratégie est plus efficace et efficiente lorsque le coût de la pénurie augmente.
Deuxièmement, nous montrons l’influence d’un système en redondance passive sur le temps
de C/D et nous comparons ces résultats avec le cas précédent (M1P1). Sur la Figure 4, nous
constatons qu’en faisant varier le coût de pénurie, le coût moyen de production augmente
pour M1P1, mais cette variation est plus faible quand le temps de C/D est fixé à une valeur
minimale. Nous pouvons gérer cette situation de façon plus fiable en intégrant la seconde
machine sous forme de redondance passive (M2P1), surtout quand le temps de C/D est fixé à
une valeur minimale.

Figure 4 Coût moyen de production/coût de pénurie

On s’intéresse également à la comparaison du coût moyen de production en faisant varier le


coût d’inventaire (Figure 5).
184

Figure 5 Coût moyen de production/coût d'inventaire

Nous constatons qu’en augmentant le coût d’inventaire, le coût moyen de production croît de
façon brutal pour M1P1, mais cette variation est plus faible si le temps du C/D est fixé à une
valeur minimale. Nous pouvons maîtriser cette variation en incorporant la seconde machine
en redondance passive.

6. Conclusion

Nos travaux permettent de confirmer qu’il est possible d’intégrer un système en redondance
passive dans le contexte manufacturier afin d’augmenter : 1) la productivité des ressources
humaines et matérielles; 2) l’efficacité et l’efficience de la production par une meilleure
gestion améliorant la disponibilité du parc machine. Cette intégration libère un espace-temps
essentiel pour minimiser les possibilités de contournement des dispositifs de protection ou
d’escamotage des procédures de C/D.
185

Références

Akella, R. and Kumar, P.R., 1986. Optimal Control of Production Rate in a Failure-Prone
Manufacturing System. IEEE Transactions on Automatic Control AC 31. pp.116-
126.
Giraud, L., Ait-Kadi, D., Ledoux, É., Pâques, J.J. et Tanchoux, S. 2008. " La maintenance,
État de la connaissance et étude exploratoire". IRSST. pp.61.
Charlot, E., Kenné, J.P. and Nadeau, S. 2006. "Optimal Production, Maintenance and
Lockout/tagout Control Policies in Manufacturing Systems". International Journal of
Production Economics 107, pp. 435-450.
Emami-Mehrgani, B., Nadeau, S and Kenné, J.P. 2011 "Lockout/tagout and operational risks
in the production control of manufacturing systems with passive redundancy".
International Journal of Production Economics, accepté.
Emami-Mehrgani, B., Nadeau, S and Kenné, J.P. 2010. "Integrating lockout/tagout with
operational risks: the passive redundancy case", Mars, Darmstadt, Allemagne,
pp.475-482.
Basile, O. et Dehombreux, P. 2002 "Reliabilitix, a Free MATLAB Toolbox for Reliability
Analysis", Faculté Polytechnique de Mons. pp.9
ANNEXE 5

TEMPS MOYEN DE CADENASSAGE/DÉCADENASSAGE : UN OUTIL


D’OPTIMISATION DES POLITIQUES DE SURVEILLANCE ET DE
MAINTENANCE

Behnam Emami-Mehrgani, Sylvie Nadeau, Jean-Pierre Kenné

Département de génie mécanique, École de technologie supérieure,

Montréal (Québec) H3C 1K3

Travail et Santé,2011,27(4),pp.22

1. Introduction

L’utilisation actuelle du C/D présente plusieurs lacunes reliées au manque d'évaluation de


leurs faisabilité et efficacité. Le risque d’erreur humaine au cours des interventions présente
une probabilité importante d’impact sous forme d’incidents ou accidents du travail chez les
maintenanciers et sur la disponibilité du système manufacturier. Le concept novateur de
«Mean Time to Logout/Tagout (MTTLT)» a été développé pour pallier à ces problèmes pour
des systèmes manufacturiers flexibles (FMS).

2. Méthodes

Le MTTLT permet l’intégration du C/D dans la gestion de la capacité de production.


Premièrement, le C/D a été modélisé pour un système manufacturier en redondance passive.
Deuxièmement, les méthodes numériques basées sur l’approche Kushner ont été utilisées afin
de déterminer la politique de production optimale du système modélisé. Finalement, les
performances de l’approche ont été évaluées à travers le modèle de simulation et le plan
d’expérience du système manufacturier étudié.
187

3. Résultats

En intégrant le MTTLT dans un système en redondance passive, le système devient moins


vulnérable aux variations des coûts de pénurie, d’inventaire et de production en satisfaisant la
demande en permanence. Cette intégration dans un système manufacturier en redondance
passive a permis de libérer un espace-temps essentiel pour minimiser les possibilités de
contournement des dispositifs de protection ou d’escamotage des procédures de C/D.

4. Conclusion

Le concept novateur MTTLT a permis d’augmenter à la fois la sécurité, afin d’éviter les
accidents graves, ainsi que l’efficacité et l’efficience de la production par une meilleure
gestion en améliorant la disponibilité du parc machine.
ANNEXE 6

UTILISATION DES DIFFÉRENTS MODÈLES POUR UN CAS RÉEL

Recueillir les données réelles du système


manufacturier (coûts, MTBF, MTTR, MTTLT)

Utiliser les différents modèles: modèle 1, modèle


2 et modèle 3

Résolutions numériques
(Matlab)

Trouver les politique optimales


U * ,U *mc ,U *mp ,U *ca

Évaluer la performance par la simulation


(Arena)

MTBF: Mean time between failures


MTTR: Mean time to repair
MTTLT: Mean time to lockouttagout
U* : Taux de production
U *mc : Taux de maintenance corrective
U*mp : Taux de maintenance préventive
U*ca : Taux de cadenassage/décadenassage (C/D)